Music playing mode intelligent switching system adapted to different scenarios

Abstract

The application discloses a music playing mode intelligent switching system suitable for different scenes and belongs to the field of artificial intelligence and ubiquitous computing. The system comprises a multi-modal perception module, which is used for collecting environment, user physiological, behavior interaction and audio content data in real time; a feature fusion coding module, which generates a joint scene feature vector through time sequence alignment, statistical feature extraction and neural network coding; a machine learning inference engine, which adopts a pre-trained multi-task learning model supporting online fine-tuning, outputs a fine-grained scene recognition result and optimal playing mode configuration parameters according to the feature vector; and a dynamic audio execution module, which analyzes the configuration parameters and realizes non-perception audio experience switching through a smooth transition controller, and simultaneously collects user feedback to form a learning closed loop. The application realizes highly personalized, context-accurate perception and smooth and seamless music playing mode full-automatic intelligent adaptation, and can be widely applied to intelligent earphones, vehicle entertainment, smart home and mobile terminals.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a music playback mode intelligent switching system adapted to different scenarios. Background Technology

[0002] In the information age, audio content has become a core digital companion in people's daily lives. From podcasts that uplift our spirits during commutes to white noise that enhances focus at work, from dynamic melodies that unleash our potential during exercise to soothing tunes that create an atmosphere at home, music and sound profoundly shape our emotions, productivity, athletic performance, and overall well-being. However, an individual's audio needs are not static but rather change dramatically and dynamically with their physical environment, physiological and psychological state, ongoing activities, and temporal context. An ideal audio experience should act like an invisible assistant, proactively, accurately, and seamlessly adapting to these changes without requiring users to frequently interrupt their current activities for manual adjustments.

[0003] Existing technologies for solving the problem of scene-adaptive audio have roughly gone through three stages of evolution, but all have significant limitations:

[0004] Phase 1: Fully Manual Mode (Static Playlists and Interface Interaction). This is the most basic solution. Users need to create and maintain different playlists in advance according to the anticipated scenarios. When using the system, users manually select the corresponding playlist through a graphical user interface. Its technical essence is a file management system with simple UI interaction. The main drawbacks include: High cognitive and operational burden: Users must interrupt their current activity, think, and perform the switching operation, disrupting the sense of immersion. Lack of dynamic adaptability: It cannot respond to changes in sub-states within the same scenario.

[0005] The second stage: based on simple rules and triggers. With the widespread use of sensors in mobile devices, automation solutions based on preset rules have emerged. For example, after a smartwatch detects running, it triggers a phone to play a workout playlist via Bluetooth; or it automatically switches to sleep mode based on the system time. Its core technology is the IF-THEN rule engine. The main drawback is its limited and coarse perception dimension: relying on only one or two sensors (such as movement or time), resulting in low accuracy and susceptibility to false triggers (e.g., fast walking being misidentified as running).

[0006] The third stage: Personalized recommendations based on collaborative filtering. This is the core of current mainstream music streaming platforms. By analyzing users' historical playback records, searches, favorites, skips, and other behaviors, personalized daily recommendations or scene-based radio programs are generated for users using collaborative filtering, matrix factorization, or deep learning models. While highly effective in content discovery, it suffers from fundamental shortcomings in real-time scene adaptation: a severe lack of context awareness; the recommendation model is essentially static and context-free.

[0007] Therefore, we have made improvements and proposed an intelligent music playback mode switching system that adapts to different scenarios. Summary of the Invention

[0008] To address the aforementioned technical challenges, this invention proposes an intelligent music playback mode switching system adapted to different scenarios. The core innovation of this invention lies in: constructing dynamic scene recognition as an online machine learning inference problem based on multimodal temporal data; defining the optimal audio playback mode as a multi-parameter joint optimization problem strongly coupled with the identified scene; and enabling the entire model to continuously evolve according to user habits through a closed-loop learning system with real-time feedback. This invention, through a meticulously designed modular architecture, realizes an end-to-end intelligent pipeline from raw sensor signals to final personalized audio output.

[0009] To achieve the above-mentioned objectives, this invention provides an intelligent music playback mode switching system adapted to different scenarios, thereby solving the aforementioned problems.

[0010] The system of this invention is a flexibly deployable software-hardware co-operation, whose core logic can run on everything from resource-constrained embedded devices (such as smart headphones) to powerful mobile or edge computing platforms (such as smartphones and in-vehicle computers), and even cloud servers. The system comprises four logically hierarchical and tightly connected core modules that together constitute an agent.

[0011] In a first aspect, the present invention provides a scene-adaptive music playback system, characterized in that it includes:

[0012] Multimodal Data Sensing Module: This module, acting as the system's sensory system, is responsible for continuously and concurrently acquiring high-dimensional, heterogeneous raw data streams from the physical world and digital interactions. Its innovation lies in the collaborative acquisition of data from multiple data sources at the levels of time synchronization, power optimization, and semantic association. This module specifically consists of four highly specialized sub-modules:

[0013] Environmental Awareness Submodule: This submodule is responsible for quantifying the user's objective physical environment. It integrates raw readings from built-in or external sensors and performs preliminary signal processing to extract meaningful features.

[0014] Acoustic Environment Analysis: Ambient sounds are continuously acquired through one or more microphones. The processing flow includes: a) Sound Pressure Level Calculation: The A-weighted equivalent continuous sound level (Leq) is calculated to assess overall loudness. b) Real-time Spectral Analysis: The audio signal is decomposed into multiple frequency bands (e.g., by 1 / 3 octave or by the psychoacoustic Bark scale) using Fast Fourier Transform (FFT) or filter banks. For each frequency band, its energy, the variance of energy over time (fluctuation), and the spectral roll-off point (frequency where the spectral energy decays to below -60 dB) are calculated. This helps to finely distinguish between steady low-frequency humming from an air conditioner, intermittent mid-frequency conversations, and continuous high-frequency keyboard clicks. c) Audio Event Detection: Optionally, a lightweight deep learning model is used to identify specific environmental sound events, such as car horns, baby cries, coffee maker sounds, and rain sounds, which are powerful scene indicators.

[0015] Optical environment perception: Ambient light intensity and color temperature are acquired through an ambient light sensor. Light intensity is related to user activity, while color is related to the sense of time and mood.

[0016] Geographic and motion environment calculation: fusing multi-source positioning and motion data. This includes: absolute position, velocity, and heading provided by Global Navigation Satellite Systems (GNSS, such as GPS and BeiDou); indoor positioning based on Wi-Fi fingerprinting, Bluetooth beacons, or ultra-wideband; and high-frequency raw data from the inertial measurement unit. Through advanced sensor fusion algorithms, the user's six degrees of freedom motion states are accurately calculated: stationary, walking, running, cycling, and riding in a vehicle.

[0017] Spatiotemporal context acquisition: Directly retrieves precise system time (year, month, day, hour, minute, second), day of the week, and holiday information from the device's operating system. It also has access to calendar applications (with user authorization) to understand the user's schedule. Time patterns are among the most powerful prior knowledge for predicting user intent.

[0018] User State Awareness Submodule: This submodule is responsible for inferring the user's internal physiological and psychological state and explicit behavioral intentions, which is key to understanding the core element of a human being.

[0019] Physiological signal monitoring: Non-invasive acquisition via biosensors. a) Photoplethysmography (PPG) sensor: Used for continuous monitoring of heart rate and heart rate variability (HRV). Time-domain and frequency-domain indices of HRV are the gold standard for assessing autonomic nervous system activity, stress levels, relaxation levels, and emotional arousal. b) Skin conductance sensor: Measures skin conductivity levels; its transient changes are closely related to emotional excitement, tension, or cognitive load. c) Body temperature sensor: Monitors changes in body surface or ear canal temperature to help determine the user's physiological cycle or environmental adaptation status.

[0020] Behavioral Interaction Log: Comprehensive monitoring of all explicit interaction events between the user and audio devices and related applications. These events constitute a timestamped sequence of behaviors, including: music player controls (play, pause, next track, previous track, fast forward, rewind), volume adjustment actions (specific increments / decrements), device physical controls (such as single click, double click, long press, swipe on headphones), recognition results of voice assistant wake words and subsequent commands, device screen unlock / lock events, and foreground application switching sequences. For example, if a user suddenly switches to an email application while playing music and remains inactive for an extended period, it may indicate that they have entered a deep working state requiring the elimination of distractions.

[0021] Attention and Context Inference: Lightweight, low-power computer vision analytics can be performed on devices equipped with a front-facing camera (with explicit user authorization and using localized edge computing to protect privacy). For example, miniature neural networks can be used to estimate the duration of a user's gaze on the screen, blink frequency, facial orientation angle, and even a rough facial expression. This information can be used to help determine whether a user is immersed in their personal device world or interacting with the physical environment or others.

[0022] Audio Content Analysis Submodule: This submodule is responsible for a deep understanding of the audio content itself, that is, what is being played and what can be played.

[0023] Metadata Acquisition and Parsing: Rich metadata is obtained from the currently playing track and candidate tracks in the media library via the operating system's media framework interface or music streaming service provider API. This includes: song title, artist, album, genre, style tag, tempo, key, loudness, release year, and music encoding.

[0024] Local Audio Signal Processing: For audio lacking metadata or requiring more refined, real-time analysis, this module performs audio feature extraction locally. Key technologies include: a) Rhythm and Beat Tracking: Estimating the beats per minute (BPM) and beat position of music using autocorrelation-based or neural network-based methods. b) Timbre and Spectral Feature Extraction: Calculating Mel-frequency cepstral coefficients and their first and second differences, which are standard features in the field of music information retrieval. c) Musical Emotion and Tonality Analysis: Determining the overall emotional color and tonality of music by analyzing chord progressions and melodic contours. d) Speech / Music Classification and Speech Activity Detection: Distinguishing between instrumental music, music with lyrics, and audio content such as podcasts / audiobooks.

[0025] Lyrics semantic analysis: With internet access and user permission, the lyrics text of a song can be obtained, and a pre-trained natural language processing model can be used for sentiment analysis and topic keyword extraction, providing richer dimensions for emotional matching of music.

[0026] User History and Preferences Module: This is a structured local database and lightweight model library used to create a digital profile of a user's long-term audio behavior. It is not only a storage for historical records, but also a probabilistic model of user habits.

[0027] Encrypted behavior log database: Securely stores all relevant events in chronological order, including records of each playback, each manual mode switch, each feedback action, and user-defined audio parameter adjustments.

[0028] Statistical preference models: These run background analytics tasks periodically (e.g., daily) to extract statistical patterns from historical data. For example, they calculate an audio feature probability distribution for each identified scenario (or location-time combination). For a gym scenario, the model might record that among the songs selected by the user, there is a 65% probability that the BPM is in the 120-140 range, a 40% probability that the genre is electronic dance music, and an average loudness 3dB higher than usual. This statistical information provides strong priors for machine learning models.

[0029] Explicit Rules and Preferences: Provides a user interface that allows users to set a few but highly impactful deterministic rules. For example: automatically play news summary podcasts from 8:00 AM to 8:30 AM every weekday; automatically activate 'Driving Mode' when a Bluetooth connection to the car is detected; and reduce the volume of all music by 20% after 11 PM. These rules have the highest priority and directly guide the system's decisions.

[0030] Scene Feature Fusion and Encoding Module: This module is the information fusion center of the system, responsible for transforming heterogeneous, heterogeneous raw data streams from different sub-modules with different sampling rates into a unified, high-quality, low-dimensional numerical representation suitable for machine learning models. Its technical challenges lie in handling the heterogeneity, temporal sequence, and noise of the data.

[0031] Data Synchronization and Time Alignment: Due to the significant differences in sampling frequencies among various sensors and event sources (IMU up to 100Hz, GPS 1Hz, heart rate 1Hz, user interaction events are asynchronous), this module first establishes a unified time reference based on the system's high-precision clock. All incoming data is timestamped. The system operates with a fixed basic processing cycle (e.g., Δt = 1 second) and analysis window length (e.g., T = 30 seconds). In each processing cycle, the module collects all available data within the past T-second time window.

[0032] Windowed feature engineering (for time series data): For each type of time series data within a window, calculate a set of robust statistical features to compress the time series into a fixed-dimensional feature vector.

[0033] For motion data (accelerometer magnitude): calculate the mean (reflecting static posture), standard deviation (reflecting overall motion intensity), zero-crossing rate (reflecting motion frequency), and the 25th, 50th, and 75th percentiles of the amplitude histogram.

[0034] For environmental acoustic data (energy in each frequency band): calculate the mean (average intensity), variance (variability), peak-to-trend ratio (dynamic range), and linear regression slope (trend) of energy values ​​in each frequency band over the past three windows.

[0035] For physiological data (heart rate): Calculate the mean, standard deviation, and root mean square (RMSSD, time-domain indicator of HRV) of the heart rate within the window. Where possible, perform a short-time Fourier transform to estimate the frequency domain component (LF, HF) of HRV.

[0036] For GPS speed: Discretize it into classification features: stationary (<1 km / h), walking (1-7 km / h), running (7-20 km / h), cycling (20-40 km / h), and vehicle (>40 km / h).

[0037] Semantic feature embedding (for categorical / symbolic data): For non-numerical data, such as geographic semantic tags (company gym), day of the week (Wednesday), currently playing genre (Synthwave), and weather conditions (sunny), embedding layers are used. These embedding layers are randomly initialized early in model training and are learned and optimized along with other parameters throughout the system training process. As a result, semantically similar categories (such as company and office) will have closer vector distances in the embedding space.

[0038] Feature normalization and dimensionality reduction: All numerical statistical features are Z-score normalized to eliminate the influence of different sensor dimensions and numerical ranges. Subsequently, principal component analysis (PCA) or autoencoders can be optionally applied for linear or nonlinear dimensionality reduction to further compress feature dimensions and remove redundancy and noise.

[0039] The fusion-encoding neural network concatenates all the features processed above (potentially containing hundreds of dimensions) into a high-dimensional feature vector. This vector is then fed into a small, but non-linearly expressive, feedforward neural network (e.g., input layer → fully connected layer (256 dimensions, ReLU activation, Dropout=0.2) → fully connected layer (128 dimensions, ReLU activation) → output layer (128 dimensions)). This network learns the complex, high-order interactions between different features and condenses the information, ultimately outputting a fixed 128-dimensional joint scene feature vector. This vector is a distributed, dense, low-dimensional representation of the environment-user-content joint state over the past T seconds, serving as the cornerstone for subsequent intelligent decision-making.

[0040] Machine Learning Inference Engine Module: This is the cognitive core and decision-making center of the system, and its intelligence level directly determines the performance of the entire system. It contains one or more machine learning models that have been pre-trained on a large scale and can be updated incrementally online (on-device). This module receives a joint scene feature vector as input and outputs two core decisions: fine-grained scene recognition results and specific playback mode control parameters.

[0041] Model Architecture Design: This invention employs a multi-task learning framework to achieve end-to-end joint inference. The model consists of a shared low-level feature extractor and multiple parallel, task-specific heads.

[0042] Shared Feature Encoder: Typically a deep neural network that receives a 128-dimensional joint scene feature vector. Its architecture can be a multilayer perceptron (MLP), or, to better capture temporal dependencies, a recurrent neural network (such as a gated recurrent unit GRU) or a one-dimensional temporal convolutional network (1D-CNN). The encoder's parameters are shared by all downstream tasks, forcing it to learn a general and robust high-level representation of the scene that is beneficial for multiple tasks.

[0043] Task 1: Fine-grained Scene Classification Head: This is a standard classifier, typically consisting of a fully connected layer and a Softmax activation function. It maps the output of a shared encoder to a probability distribution of a predefined, fine-grained set of scene categories. The design of the scene categories is crucial, balancing generality and personalization. An example set might include: Commuting - Walking, Commuting - Driving - Highway, Commuting - Driving - Traffic Jam, Work - Deep Focus, Work - Creative Thinking, Work - Routine Tasks, Exercise - Aerobic Running, Exercise - Strength Training, Studying - Reading, Studying - Attending Lectures, Leisure - Social Gathering, Leisure - Relaxing Alone, Home - Cooking, Home - Cleaning, Bedtime Preparation, Sleeping, Unknown. The output probabilities are not only used for the final decision, but their maximum value is also a key threshold for determining whether to trigger a switch.

[0044] Task 2: Playback Mode Parameter Prediction Head: This is a multi-output regression head (for continuous parameters) or classification head (for discrete parameters). It directly predicts the specific configuration of the optimal playback mode, forming a playback mode configuration vector. This vector is a blueprint for the system execution level and may contain:

[0045] Content control parameters: an ID pointing to a specific playlist or radio station; or a target point embedding vector in a music feature space (such as a space spanned by genre, rhythm, and sentiment values) for real-time retrieval of similar songs.

[0046] Audio signal processing parameters:

[0047] Target average loudness (LUFS).

[0048] Gain (dB) of the multi-band parametric equalizer at each center frequency (e.g., 31Hz, 62Hz, 125Hz, 250Hz, 500Hz, 1kHz, 2kHz, 4kHz, 8kHz, 16kHz).

[0049] Threshold, ratio, start-up time, and release time of the dynamic range compressor.

[0050] Spatial audio rendering parameters, such as virtual sound field width, surround sound intensity, and sound source height simulation.

[0051] Equipment hardware control parameters:

[0052] Intensity level of active noise cancellation / ambient sound pass-through (0%-100%).

[0053] Enable dialogue enhancement mode.

[0054] Audio codec preferences (such as prioritizing low-latency aptX Adaptive or high-fidelity LDAC).

[0055] Playback strategy parameters:

[0056] Playback order: sequential, random, and intelligent interleaving based on similarity.

[0057] Allow system notification sounds to interrupt music.

[0058] The default duration of the cross-fade-in and fade-out between songs.

[0059] Task 3 (Optional): Context-Aware Instant Song Scoring Head: This is a variant of a neural collaborative filtering or deep ranking model. It receives two inputs: a user's current context vector output by a shared encoder, and a song feature vector (which can be metadata embeddings or audio fingerprint vectors) from the audio content analysis submodule. Through an interactive network (such as an inner product or multilayer perceptron), it calculates the user's instant preference score for the song in the current context. This score is used to reorder the candidate song list, ensuring that the next song pushed to the user not only conforms to the scene pattern but also highly matches the user's instantaneous state.

[0060] Large-scale pre-training in the cloud: While ensuring data privacy and security, massive amounts of anonymized multimodal time-series data are collected on cloud servers. This data is correlated with the user's final music interaction behavior (which can serve as indirect labels). Using this data, the aforementioned multi-task model is trained in an end-to-end manner. The loss function is a weighted sum of the losses for each task: `L-total = α L-scene-classification + β L-parameter-regression + γ L-song-ranking`.

[0061] Federated Learning and Privacy Protection: To protect user data privacy and achieve global model improvement, this invention supports a federated learning framework. The initial model is a pre-trained version in the cloud. User devices participating in federated learning calculate model updates (gradients) locally using their own data and upload encrypted gradient updates to the cloud server. The server aggregates updates from a large number of devices, generates a new global model, and then distributes it to each device. The user's original data never leaves the device.

[0062] On-device local fine-tuning and personalization: When a user uses the system for the first time or when the model is updated periodically, the system uses locally stored, encrypted user history behavior data to fine-tune the global model several times. This process allows the model to quickly adapt to the unique habits of a specific user (for example, the user's defined focus scenario may prefer post-rock music to plain white noise).

[0063] Online reinforcement learning and real-time adaptation: During system operation, each decision (scene recognition + mode switching) triggers immediate feedback (positive / negative) from the user. This essentially constitutes a reinforcement learning environment. The system stores each interaction as a sample of a four-tuple `(state, action, reward, next-state)` in a fixed-size experience replay buffer. Periodically, a batch of samples is sampled from the buffer, and the model (especially the policy network part) is updated online in small increments using online policy gradients or variants of Q-learning. This allows the system to quickly respond to short-term changes in user preferences or adapt to entirely new, undefined scenarios.

[0064] Dynamic audio engine and switching execution module: This is the system's motor nervous system and effector, responsible for translating abstract intelligent decisions into concrete changes in the user's audio experience that are perceptible, and ensuring that the entire change process is smooth, natural, and even imperceptible.

[0065] The policy parser and instruction dispatcher receive the playback mode configuration vector from the inference engine and compile it into a series of specific instructions that can be understood by the underlying hardware and software APIs. For example, it converts EQ: Low frequency +3dB, Mid frequency -2dB, High frequency +1dB into a coefficient array sent to the Audio Digital Signal Processor (DSP); and ANC: 70% into a specific HCI command sent to the Bluetooth headset chip.

[0066] Smooth Transition Controller: This controller is key to a superior user experience. Its goal is to eliminate any auditory breaks or jumps.

[0067] State Machine Management: The controller maintains an internal state machine with states including: IDLE (stable), PLANNING-TRANSITION (new decision received), IN-TRANSITION (transitioning), and EVALUATING (observing after transition). The state machine ensures the rigor of the transition logic and implements anti-jitter functionality to prevent high-frequency oscillations between high and low confidence levels caused by sensor noise.

[0068] Parameter interpolation and easing: For all continuously changing parameters (volume, EQ gain, soundstage width, ANC intensity), the controller generates a time-to-value curve for each parameter, from the current value to the target value. The curve is not a simple linear change, but rather employs an easing function that conforms to human perception, such as ease-in-out-cubic, making the change smoother at the beginning and end, and faster in the middle. The transition duration is dynamically adjusted based on the parameter type and the magnitude of the change, typically between 0.3 and 5 seconds.

[0069] Intelligent audio content transition:

[0070] Crossfade-in / fade-out: When switching between two songs, instead of simply overlaying a fade-in / fade-out at the end of one song and the beginning of another, the controller analyzes the beat grid of both songs. If possible, it aligns the switch point to the strong beats of both songs and completes the fade-out within a preset multiple of beats (such as 4 or 8 beats), achieving a seamless DJ mix feel.

[0071] Tonal Matching (Optional): On high-end devices, the tonalities of two songs can be analyzed in real time, and a slight real-time pitch shift can be applied during the transition to make the two songs more tonal harmonious.

[0072] Parallel transition management: The controller coordinates the transition process of all parameters, ensuring they are synchronized in time or performed in a specific order to avoid creating discordant auditory effects. For example, it might start by reducing the volume of an older song, then apply a new EQ curve, and finally initiate the fade-in of a new song.

[0073] Real-time feedback acquisition and quantifier: Within a critical evaluation window (e.g., 60 seconds) after a mode switch, this unit highly sensitively monitors all user interactions. It defines a clear and quantifiable set of reward signals:

[0074] Positive feedback (+reward): Play the switched song completely; click "like" or "favorite" during playback; increase the volume (within a reasonable range); do not perform any operation for a long time (meaning accept the current state).

[0075] Negative feedback (-reward): Skip the song for a short period of time (e.g., 15 seconds) after it starts playing; manually switch back to the previous mode or song; manually make significant adjustments to parameters that the system just set automatically (e.g., immediately turn off the noise reduction that was just turned on); directly pause playback and exit the application.

[0076] These reward signals are quantified into a scalar value rt, which, together with the previous decision data, constitutes a reinforcement learning sample.

[0077] Safety barriers and exception handlers: To ensure the reliability and security of the system, this module integrates a series of mandatory rules:

[0078] Driving safety rules: When the system is certain that the user is driving (via GPS speed, Bluetooth connection to the vehicle, etc.), the cinema mode that completely blocks ambient sound is prohibited from being enabled, and the notification sound channel must be kept open.

[0079] Hearing protection rules: Based on time-weighted average sound pressure level calculation, to prevent hearing damage from prolonged high-volume playback, automatically implement volume limits.

[0080] Power-sensitive policy: Automatically disables high-power complex audio processing and continuous high-frequency sensor sampling when the device has low battery.

[0081] User veto right: Users can temporarily freeze the system's automatic switching function for a period of time by using a shortcut (such as telling the headset to keep the current mode).

[0082] Secondly, this invention provides a method for intelligent switching of music playback modes based on the aforementioned system. This method essentially describes the complete cognition-action cycle of the aforementioned intelligent agent within a processing cycle. Its characteristic is that the method operates on a computing device including a processor, memory, and multiple sensors, and includes the following steps:

[0083] S1: System Initialization. Load the pre-trained machine learning model into memory, initialize the data streams of each sensor, the feature buffer, and the state of the smooth transition controller. Load the user's historical preference model from the local database.

[0084] S2: Enter the main perception-decision-execution loop. The loop runs at a fixed frequency (e.g., 1Hz) or in an event-driven manner.

[0085] S2.1: Parallel acquisition of multimodal data. The environmental awareness, user status awareness, and audio content analysis submodules concurrently acquire the latest data packets and timestamp them.

[0086] S2.2: Scene Feature Fusion and Encoding. For the current processing time t, the scene feature fusion module collects all data within the time window [tT, t]. It performs feature engineering, semantic embedding, normalization, and other operations, and finally generates the joint scene feature vector Ct for the current time through a fusion encoding neural network.

[0087] S2.3: Machine Learning Inference. Input Ct into the machine learning inference engine module. The model performs forward propagation and outputs: a) a fine-grained scene classification probability distribution Pt = {p-1, p-2, ..., pN}; b) a recommended playback mode configuration vector Mt; c) (optionally) a context-reordered list of candidate songs Lt.

[0088] S3: Decision evaluation and switching trigger.

[0089] S3.1: Find the scene label S-new = argmax(Pt) with the highest probability from Pt, and its corresponding probability is confidence.

[0090] S3.2: Get the currently executing active scenario S-current and active mode M-current.

[0091] S3.3: Apply the switching decision function F-switch(S-current, S-new, confidence, hysteresis). This function not only compares S-current and S-new, but also considers whether the confidence exceeds a threshold θ (e.g., 0.7), and introduces hysteresis logic to prevent oscillations near scene boundaries. For example, a switching decision is only made when S-new != S-current, confidence > θ, and S-new has been continuously identified for more than K cycles.

[0092] S4: Smooth execution switching.

[0093] S4.1: The smooth transition controller enters the PLANNING-TRANSITION state. It analyzes the difference between M-current and Mt, and formulates a detailed transition plan (target value, easing curve, duration) for each parameter that needs to be changed.

[0094] S4.2: Notify the audio subsystem to prepare new audio content (select the highest priority song from Lt, and may preload it).

[0095] S4.3: The controller enters the IN-TRANSITION state and concurrently initiates all planned transition processes. During the transition, it updates the parameter interpolation results at a high frequency (such as the audio sampling rate or display refresh rate) and applies them synchronously to the audio rendering pipeline.

[0096] S4.4: After all transitions are complete, the controller enters the EVALUATING state. The system updates S-current to S-new and M-current to Mt.

[0097] S5: Feedback collection and experience storage.

[0098] S5.1: During the duration of the EVALUATING state (e.g., 60 seconds), the real-time feedback collector monitors user interaction and ultimately quantifies a reward value r.

[0099] S5.2: Encapsulate the experience from this loop into a tuple (Ct, S-new, Mt, r, C-{t+60}), where C-{t+60} is the new scene feature vector 60 seconds after the switch (representing the subsequent state). Store this experience sample in the local experience replay buffer.

[0100] S6: Continuous learning and model updates.

[0101] S6.1: Periodically check (e.g., every 100 new samples collected, or at midnight each day) and start a background learning task if the device is idle and charging.

[0102] S6.2: Randomly sample a mini-batch of empirical data from the empirical replay buffer.

[0103] S6.3: Using this data, perform an incremental update on the local machine learning model using stochastic gradient descent or its variants. The update can be applied to the entire model or only to the last few layers (a concept from transfer learning) to quickly adapt to new knowledge.

[0104] S6.4: Under the condition of satisfying privacy and security policies, the local model update (gradient) is encrypted and uploaded to the cloud to participate in federated learning aggregation.

[0105] S7: Return to step S2.1 and continue the loop for the next processing cycle.

[0106] The beneficial effects of this invention are that, compared with the prior art, it has the following significant and verifiable advantages:

[0107] 1. By systematically integrating six dimensions of information—physical environment, user status, audio content itself, and long-term history—on consumer-grade audio devices, a dynamically updated user context profile far exceeding existing technologies has been constructed, providing an unprecedented data foundation for accurate scene understanding.

[0108] 2. An innovative multi-task learning model architecture integrates multiple highly related tasks, such as scene recognition, playback parameter prediction, and song recommendation, under a unified framework for joint optimization. This avoids the problems of error accumulation and feature inconsistency in traditional serial pipelines, enabling the discovery of deeper correlations in the data and achieving better and more collaborative global decision-making.

[0109] 3. A unique, smooth transition controller based on easing functions and intelligent audio analysis transforms mode switching from unpleasant auditory abrupt stops into elegant, perceptually consistent auditory gradations. It encompasses comprehensive smoothing from macro-level content to micro-level sound effect parameters, truly achieving technological invisibility and enhancing the smoothness and professionalism of the user experience.

[0110] 4. By integrating a hybrid learning paradigm that combines federated learning, on-device fine-tuning, and online reinforcement learning, the system can continuously learn from users' real-time interactions while strictly protecting user data privacy. This allows the system to not only adapt to long-term habits but also capture short-term preference changes.

[0111] 5. A clear modular design allows for flexible deployment and customization of the system. From smartwatches with only a basic IMU and heart rate sensor to flagship smartphones with rich sensors and powerful smart cars, different levels of intelligent functions can be adapted by enabling or disabling corresponding sub-modules. The system API design allows for integration with various audio hardware and streaming services. Attached Figure Description

[0112] Figure 1 The internal flowchart of the multimodal data perception module of the intelligent music playback mode switching system adapted to different scenarios provided by the present invention.

[0113] Figure 2 The internal flowchart of the scene feature fusion and encoding module of the intelligent music playback mode switching system adapted to different scenarios provided by the present invention;

[0114] Figure 3 Flowchart of the machine learning inference engine module of the intelligent music playback mode switching system adapted to different scenarios provided by the present invention.

[0115] Figure 4 The internal flowchart of the dynamic audio engine and switching execution module of the intelligent music playback mode switching system adapted to different scenarios provided by the present invention;

[0116] Figure 5 The flowchart shows the internal process of the incremental learning and model update module of the intelligent music playback mode switching system adapted to different scenarios provided by this invention. Detailed Implementation

[0117] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0118] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0119] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the present invention can be combined with each other.

[0120] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0121] Example 1: Integrated into high-end true wireless smart earbuds and their accompanying mobile applications

[0122] This embodiment describes in detail the specific implementation of the present invention in a true wireless smart earphone ecosystem with advanced sensors and collaborative processing capabilities.

[0123] 1. Hardware platform configuration:

[0124] Smart earphones (left / right):

[0125] Main controller: Dual-core low-power Bluetooth audio SoC (such as Qualcomm QCC5141 series), with built-in high-performance DSP and neural network accelerator (NPU).

[0126] sensor:

[0127] Inward and outward beamforming microphone arrays (2-3 of each).

[0128] Six-axis IMU (accelerometer + gyroscope).

[0129] Wear a detection sensor (optical or capacitive).

[0130] Touch-sensitive area.

[0131] PPG heart rate sensor (located at the ear canal contact point).

[0132] Audio: Supports adaptive active noise cancellation, transparency mode, and high-definition audio codecs.

[0133] Smartphones: High-performance mobile platforms (such as Apple A series or Qualcomm Snapdragon 8 series) running a matching smart audio center app.

[0134] 2. Software Architecture and Responsibility Allocation (Collaborative Computing):

[0135] Considering the headphone's battery life and computing power, a computing architecture combining edge-cloud collaboration and mobile phone-headphone collaboration is adopted.

[0136] Headphone end (ultra-low power consumption, always-on distributor):

[0137] Running lightweight firmware, it is responsible for continuously sampling PPG and IMU raw data with extremely low power consumption (<1mW).

[0138] Perform preliminary, fixed signal processing: such as calculating real-time motion energy values, estimating heart rate, and performing simple voice activity detection (for wake words).

[0139] The processed feature data (not the raw audio stream, to save bandwidth and power consumption) is packaged and sent to the mobile app via Bluetooth Low Energy (BLE) channel at a frequency of 1-5Hz. The raw audio stream is used only for ANC and calls.

[0140] Mobile App (Main Decision and Computing Center):

[0141] Multimodal data sensing module:

[0142] Receive BLE data packets from the headset.

[0143] The iOSCoreMotion / AndroidSensorManager API is used to obtain high-precision IMU data from the phone itself (for fusion with headphone data to improve motion recognition accuracy) and barometer data (to assist in floor determination).

[0144] Obtain GPS location, speed, and heading using CLLocationManager.

[0145] Access system media playback information to obtain metadata for the currently playing song. For unsupported applications, enable the bypass analysis function of the audio content analysis submodule (user authorization is required to access system audio).

[0146] The user history and preferences module is implemented using an encrypted SQLite database.

[0147] Scene feature fusion and encoding module: Using a 30-second analysis window and a 5-second sliding step (making a decision every 5 seconds), it fuses all data from the past 30 seconds. The generated feature dimension is approximately 180, including: mean heart rate and RMSSD, exercise energy percentile, environmental noise Leq and variability in 6 key frequency bands, GPS speed classification, Wi-Fi SSID (as location semantics), time features (hours, minutes, whether it's a weekday), current song genre and BPM, etc. A two-layer MLP (180->128->64) is used for encoding, outputting a 64-dimensional joint scene feature vector.

[0148] Machine Learning Inference Engine Module: Deploys a lightweight multi-task MLP model on the mobile device. The model has been pre-trained in the cloud using anonymized data from tens of thousands of users. Upon first app launch, it performs rapid fine-tuning using the user's local history from the past week (if applicable) (takes approximately 1-2 minutes). The model outputs probabilities for 12 scenes and a 15-dimensional playback mode vector.

[0149] Dynamic audio engine and switching execution module:

[0150] Control music playback using the MediaSession API.

[0151] Send EQ parameters and ANC control commands via standard Bluetooth A2DP / AVRCP commands or proprietary SDKs provided by the headphone manufacturer.

[0152] The smooth transition controller manages all transitions, with transition times typically ranging from 1 to 3 seconds.

[0153] The feedback collector monitors changes in playback status.

[0154] 3. Typical User Journey and System Response:

[0155] 08:00, at home: The user gets up and puts on headphones. The system detects: heart rate slowly rising from the resting sleep value, slight movement (getting dressed), quiet environment, and 8:00 AM. The initial scenario is "Home_Early Morning". The system plays a playlist of light morning wake-up music that the user often listens to, with ANC off and the volume at medium.

[0156] 08:15, Walking to the subway station: The user exits the building. The system detects: regular walking pattern, city street noise (mainly low to medium frequency), steadily increasing heart rate, and GPS-based walking speed. The scene switches to Commuting - Walking. The system smoothly transitions to the Commuting podcast list and adjusts ANC to Transparency Mode - Mild Noise Reduction to allow the user to perceive ambient sounds and ensure safety.

[0157] At 08:25, upon entering the subway car: the system detected a drastic change: the movement state abruptly changed from walking to standing still (relative to the subway), the ambient noise spectrum changed to a strong, stable low-frequency roar (typical subway sound), and the GPS signal was lost / drifted. The feature fusion module captured this dramatic change. The inference engine determined with >90% confidence that the scene had switched to commuting_subway. The dynamic engine executed: 1) enhancing ANC to maximum noise reduction within 2 seconds; 2) triggering pre-caching of the next 3 podcasts due to potential network instability; 3) fine-tuning the EQ, slightly boosting mid-to-high frequencies to enhance speech clarity and counteract the masking effect of low-frequency noise.

[0158] 09:00, arrives at office workstation: User sits down. System detects: prolonged stillness, connection to company Wi-Fi, ambient noise stabilizes to air conditioning hum, heart rate drops to near resting level, phone enters charging mode. Scene switches to Work - Deep Focus. System: 1) Switches to a focus-focused playlist (primarily instrumental electronic ambient music and natural sounds); 2) Sets ANC to full noise cancellation; 3) Applyes an EQ preset that reduces the 250Hz-1kHz vocal frequency range to further eliminate potential speech interference.

[0159] At 11:30, repetitive data processing began: the user started intensively working with Excel. The system detected an increased cognitive load by monitoring the application switching sequence (Music App → Excel → Email → Excel, with no music control for an extended period) and a slight decrease in HRV. The system automatically fine-tuned its mode: selecting slower, more abstract tracks from the playlist and keeping ANC at its highest level. The entire process was completed in the background without any pop-up notifications.

[0160] At 18:00, heading to the gym: The user gets up, and the GPS indicates they are moving towards the gym. The system pre-loads a strength training playlist. Upon entering the gym, it detects a strong IMU signal (weightlifting), high ambient noise, and a rapid increase in heart rate, switching the scene to Exercise - Strength Training. The system plays high-energy electronic music, applies a powerful bass EQ, and switches the ANC to transparency mode (allowing the user to hear coaching instructions or ambient sounds).

[0161] During a subway commute, a user manually skipped a piece of instrumental music automatically selected by the system and instead chose an audio podcast. This skipping behavior was recorded as negative feedback. That evening, while the phone was charging, background learning was initiated, using a batch of data containing that sample to fine-tune the model. A few days later, in the same scenario, the system's probability of recommending podcasts significantly increased.

[0162] Example 2: Deployment in the cockpit of an intelligent electric vehicle This embodiment describes the implementation of the present invention in an intelligent car cockpit with powerful computing capabilities, abundant sensors, and stringent safety requirements.

[0164] System features and expansions:

[0165] Central computing platform: The system operates as a high-priority service on the in-vehicle infotainment system (IVI) or domain controller.

[0166] Extended multimodal sensing:

[0167] Vehicle bus data: Real-time data such as vehicle speed, engine speed, accelerator / brake pedal position, steering angle, driving mode (Comfort / Sport / Eco), door / window status, outside temperature, and windshield wiper status are acquired via Controller Area Network (CANFD) or Ethernet. This data is crucial for assessing driving behavior and the in-vehicle environment.

[0168] In-cabin sensing system:

[0169] Driver status monitoring: A dedicated infrared camera (local processing) analyzes the driver's facial features in real time and outputs abstract indicators such as attention score, drowsiness level, head orientation, and gaze direction. The raw image data is not stored or uploaded.

[0170] Occupant perception: A ceiling microphone array is used for sound source localization and voice detection to determine the number of passengers and the level of conversation. An optional TOF sensor can be used to obtain the approximate posture of the occupants.

[0171] In-cabin optical sensors: monitor overall illumination and color temperature.

[0172] Specific scenario definitions: deeply integrated with driving tasks. Examples include: highway cruising (smooth), city traffic congestion (frequent stops and starts), mountain driving (aggressive), long-distance night driving, parking and charging (entertainment), in-car meetings, and family trips (children's entertainment).

[0173] Complex playback mode controls:

[0174] Advanced sound field management: Supports independent audio zoning for passenger areas. For example, in family travel mode, navigation prompts and music can be played for parents in the front row, while cartoon audio can be played for children in the back row, and acoustic isolation between the two areas can be achieved using reverse sound wave cancellation technology.

[0175] Safe Audio Mixing and Prioritization: The system must intelligently manage audio priorities. Navigation prompts, collision warnings, and incoming call ringtones have the highest priority. When these events occur, the music is skipped—the volume is quickly reduced and then smoothly restored after the prompt ends. The system can also analyze external microphone signals, identify emergency vehicle sirens, and automatically insert warning sounds into the corresponding directional (left / right) audio channels.

[0176] Dynamic integration with vehicle performance: In Sport driving mode, not only can you play exciting music, but you can also subtly couple the vehicle's actual acceleration G-force, engine speed (for gasoline vehicles) or motor whine (for electric vehicles) with the music's rhythm or bass frequencies to enhance immersion. Simultaneously, it strengthens the vehicle's active road noise cancellation function.

[0177] Machine learning models: Due to the powerful computing capabilities of vehicles, more complex models, such as Transformer encoders, can be used to process longer historical context sequences (e.g., data from the past 5 minutes) to better predict the continuation and changes in driving scenarios. These models require specialized training to understand driving safety-related features and constraints.

[0178] Scenario: A long weekend road trip, the vehicle is fully loaded.

[0179] Initial state: The vehicle has just entered the highway and is cruising smoothly. The system detects: vehicle speed is stable at 110km / h, driving mode is set to Comfort, the DSM (Driver Monitoring System) shows the driver is attentive, and passengers are conversing in hushed tones. The scenario is determined to be highway cruising_smooth. The system plays the driver's preferred long-distance driving music playlist, with the sound focused throughout the vehicle at a moderate volume, and activates active noise cancellation (primarily to eliminate wind and tire noise).

[0180] Change 1: Entering complex mountain roads with more curves. The system detects: frequent changes in steering angle, slight braking, and the driving mode may automatically or manually switch to Sport. DSM may detect increased driver focus. The scenario is fine-tuned to mountain driving. The system may automatically select more upbeat and energizing music and slightly increase the volume.

[0181] Change 2: The child in the back seat starts crying. The cabin microphone detects a high-frequency crying sound, and the sound source is located in the back seat. The system recognizes the occupant's discomfort. It can automatically perform one or more of the following actions: 1) Increase the volume of the rear entertainment system (play content that the child likes); 2) Play masking, soothing white noise for the parents through the front headrest speakers; 3) Suggest to the driver whether they need to find the nearest service area to rest.

[0182] Safety First: Throughout the process, any advanced audio effects (such as strong sound field isolation) will not be activated if the driver is visibly distracted or in complex road conditions. Navigation commands always have absolute priority over interrupting and dodging music.

[0183] Example 3: Smart Home Audio System with Multi-Device Collaboration

[0184] This embodiment describes the implementation of the present invention in a home Internet of Things ecosystem consisting of multiple smart speakers, smart displays, and wearable devices.

[0185] Distributed sensing network: Each smart device in the home (speaker, TV, watch, door lock, lighting) is a sensing node. Data is merged in the home gateway or home cloud. All data undergoes strict de-identification processing before merging, using only device IDs and abstract characteristics, and is not associated with any specific individual unless explicitly authorized.

[0186] User identification and association: Users and data are associated through multiple methods while maintaining privacy: a) Voiceprint recognition (processed locally on the device, outputting only the user ID); b) Bluetooth ID of the user's wearable device; c) Manual login on a specific device (such as a personal mobile phone). The system maintains an independent preference model for each family member.

[0187] Family-level scenario definition: waking up the whole family in the morning, quiet during weekdays, dinner preparation time, home theater time, bedtime reading, and nighttime security.

[0188] Multi-device audio orchestration: As a home audio conductor, the system can orchestrate multiple speakers to play the same content synchronously (whole-house audio), play different content (zone audio), or achieve audio follow-up—when a user moves from the living room to the bedroom, the music source smoothly shifts from one speaker to another.

[0189] The family wake-up scenario in the early morning: Based on the information of the first member who wakes up (detected by their wristband), the system begins playing soft, gradually increasing birdsong and stream sounds on the smart speaker in their bedroom, while simultaneously and slowly brightening the smart lights in that room (simulating sunrise). 30 minutes later, the system determines that other members may also be waking up, and begins playing a morning news briefing on the speakers in the kitchen and living room.

[0190] Personal reading time scenario: The system uses indoor sensors (such as human detection from a camera in privacy mode) and the user's mobile phone status to determine if a member is sitting alone in the study for an extended period of time. The system then turns the volume of the study's speakers down to a very low level, plays background music unique to that member (such as classical music or rain sounds), and automatically pauses or lowers the volume of background music in other public areas of the home.

[0191] Friend gathering scenario: The system detects multiple sound sources, active conversations, evening hours, and the possibility of more mobile devices connecting to the home Wi-Fi. The system automatically switches to party mode: plays upbeat music, arranges multiple speakers into a stereo or surround sound array to enhance the atmosphere, and may synchronize with smart lighting to switch to a dynamic color mode.

[0192] In this invention, unless otherwise explicitly specified and limited, terms such as installation, connection, and fixation should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances. Obviously, the embodiments described above are only some embodiments of this invention, not all embodiments. The accompanying drawings show preferred embodiments of this invention, but do not limit the patent scope of this invention. This invention can be implemented in many different forms; on the contrary, the purpose of providing these embodiments is to make the disclosure of this invention more thorough and complete. Although the invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the patent protection scope of this invention.

Claims

1. A music playback mode intelligent switching system adaptable to different scenarios, characterized in that: include: The multimodal data perception module is used to collect environmental perception data, user status perception data, audio content analysis data, and user historical preference data in parallel. The scene feature fusion and encoding module is connected to the multimodal data perception module and is used to perform temporal alignment, feature engineering and deep encoding on the collected heterogeneous data, and output a unified joint scene feature vector. The machine learning inference engine module is connected to the scene feature fusion and encoding module. It has at least one pre-trained multi-task learning model built in, which is used to receive the joint scene feature vector and jointly output a fine-grained scene classification probability distribution and a corresponding playback mode configuration vector. The playback mode configuration vector includes audio content preferences, audio processing parameters and device control instructions. The dynamic audio engine and switching execution module are connected to the machine learning inference engine module. They are used to parse the playback mode configuration vector, generate low-level control commands, manage the switching process through a smooth transition controller, and collect user interaction feedback. The user interaction feedback, together with the corresponding joint scene feature vector and playback mode configuration vector, constitutes the training samples, which are used to incrementally learn the model in the machine learning inference engine module.

2. The intelligent music playback mode switching system adapting to different scenarios according to claim 1, characterized in that, The multimodal data sensing module includes: The environmental perception submodule is used to acquire environmental acoustic spectrum characteristics, optical characteristics, geographic location semantic information, and user physical motion patterns through the device's physical sensors. The user state awareness submodule is used to acquire the user's physiological signal characteristics, explicit interaction behavior sequences with audio devices, and attention state inferred based on sensor fusion. The audio content analysis submodule is used to analyze the acoustic features, metadata tags, and lyric sentiment semantics of the currently playing and candidate audio streams in real time. The User History and Preferences submodule is used to store long-term user behavior logs and build a personalized scenario-audio feature probability mapping model based on statistical learning.

3. The intelligent music playback mode switching system adapting to different scenarios according to claim 2, characterized in that, The environmental perception submodule analyzes the acoustic environment by calculating the equivalent continuous A-weighted sound level, spectral roll-off point, zero-crossing rate, and time-domain envelope fluctuation coefficient of the ambient sound in multiple preset frequency bands. The user state perception submodule analyzes physiological signals by calculating the time-domain standard deviation of heart rate, the root mean square of the difference between adjacent NN intervals, and the low-frequency to high-frequency power ratio of heart rate variability, which are used to quantify the user's stress, relaxation, or excitement level.

4. The intelligent music playback mode switching system adapting to different scenarios according to claim 1, characterized in that, The operations performed by the scene feature fusion and encoding module include: Establish a unified time reference and resample and align sensor data at different sampling rates using a sliding time window as the unit; For the time series data within the window, extract statistical features including mean, variance, skewness, kurtosis, zero-crossing rate, frequency domain energy percentage, and short-term trend slope; Semantic category data is converted into dense vectors using an embedding layer; After concatenating all features, the input is a feedforward neural network containing at least one fully connected layer and an activation function for nonlinear fusion and dimensionality reduction, and the output is the joint scene feature vector with fixed dimensions.

5. The intelligent music playback mode switching system adapting to different scenarios according to claim 1, characterized in that, The multi-task learning model in the machine learning inference engine module includes: A shared feature encoder, consisting of multiple fully connected layers, is used to extract high-level abstract representations from the joint scene feature vector; A scene classification head is connected to a shared feature encoder, which uses the Softmax activation function to output the fine-grained scene classification probability distribution; A mode parameter prediction head is connected to a shared feature encoder and uses a multi-output regression structure to directly predict the values ​​of each continuous parameter in the playback mode configuration vector. An optional content scoring head receives the output of the shared feature encoder as a user context vector and interacts with the song feature vector from the audio content analysis submodule to output a context-dependent instant song preference score.

6. The intelligent music playback mode switching system adapting to different scenarios according to claim 5, characterized in that, The playback mode configuration vector includes at least four of the following parameters: the embedding vector identifier of the target music genre or playlist, the upper and lower limits of the target tempo range, the target average loudness value, the gain value of the multi-band equalizer at each center frequency, the threshold and ratio of the dynamic range compressor, the intensity level of active noise cancellation / ambient sound pass-through, the virtual sound field width parameter, the playback order strategy identifier, and the crossfade-in and crossfade-out duration.

7. The intelligent music playback mode switching system adapting to different scenarios according to claim 1, characterized in that, The dynamic audio engine and the smooth transition controller in the switching execution module are configured as follows: When it is determined that a playback mode needs to be switched, a transition time curve from the current value to the target value is generated for all continuous parameters to be changed. The curve adopts a linear, quadratic, or sinusoidal easing function. During the transition time, the parameter interpolation results are gradually applied to the audio processing pipeline at a frequency higher than that perceived by the human ear. When switching audio content, based on the analysis of the rhythm and key of the two songs, intelligent cross-fade-in and fade-out are performed, and automatic beat alignment can be optionally performed.

8. The intelligent music playback mode switching system adapting to different scenarios according to claim 1, characterized in that, The system adopts a personalized learning paradigm that combines federated learning with local online learning: The model of the machine learning inference engine module is first pre-trained in the cloud using an anonymized dataset; After the model is deployed to the user's device, it is fine-tuned using locally stored historical user data; Feedback samples generated during system operation are periodically uploaded to the cloud server in a secure aggregation manner to update the global model, or to perform small-batch online gradient descent directly on the local device to update the local model parameters.

9. A method for intelligent switching of music playback modes based on the system according to any one of claims 1-8, characterized in that, Includes the following steps: S1: Continuously and in parallel acquire multimodal sensing data streams; S2: Perform time window segmentation, feature extraction, and fusion encoding on the perceived data at a fixed period to generate a joint scene feature vector; S3: Input the joint scene feature vector into the machine learning model for inference to obtain the scene classification result and playback mode configuration; S4: Based on the comparison between the scene classification confidence and the current state, determine whether a mode switch is triggered. If triggered, start the smooth transition process to gradually change the audio system state from the current mode to the target mode. S5: During the observation period after the switch, collect user interaction behavior as immediate feedback; S6: Encapsulate the feature vector, execution mode, and feedback of this decision into experience samples, store them in a local buffer, and use them periodically for incremental learning of the model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in claim 9.

Images