Voice recognition system driven by AI
By designing an AI-driven speech recognition system, using advanced preprocessing technology and deep learning algorithms, the problems of low recognition rate and noise sensitivity of traditional speech recognition systems are solved, achieving higher recognition accuracy and flexibility.
Patent Information
- Application Number
- CN202411622909.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-13
- Publication Date
- 2025-02-18
- Estimated Expiration
- Not applicable · inactive patent
Smart Images
Figure CN119479647A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI-driven speech recognition system. Background Art
[0002] The AI speech recognition system is an innovative system based on artificial intelligence technology that can convert human speech into recognizable text or instructions, enabling computers to understand and respond to human language. The system relies on advanced acoustic and language models, as well as the support of deep learning algorithms, and can accurately identify various speech features, including pronunciation, intonation, accent, etc.
[0003] With the rapid development of artificial intelligence technology, speech recognition technology has become one of the important ways of human-computer interaction. Traditional speech recognition systems have problems such as low recognition rate and noise sensitivity, which are difficult to meet the needs of practical applications. Therefore, it is of great significance to develop an efficient and accurate AI-driven speech recognition system. Summary of the invention
[0004] In view of the deficiencies of the prior art, the present invention provides an AI-driven speech recognition system that solves the problems raised in the background technology.
[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: an AI-driven speech recognition system, comprising an audio input module, a preprocessing module, a feature extraction module, an acoustic model module, a language model module, and an output module, wherein the audio input module is used to capture and convert sound signals into digital signals, the preprocessing module cleans and formats the captured audio signals, the feature extraction module extracts features that are helpful for speech recognition from the preprocessed audio signals, the acoustic model module is used to map audio features to phonemes or phoneme combinations, the language model module is used to represent vocabulary, grammar, and semantic probability distributions, and the output module converts the recognized text sequences into a readable form.
[0006] Preferably, the output end of the audio input module is electrically connected to the input end of the preprocessing module, the output end of the preprocessing module is electrically connected to the input end of the feature extraction module, the output end of the feature extraction module is electrically connected to the input end of the acoustic model module, the output end of the acoustic model module is electrically connected to the input end of the language model module, and the output end of the language model module is electrically connected to the input end of the output module.
[0007] Preferably, the audio input module includes a sound sensor unit and a signal conversion unit, the preprocessing module includes a noise reduction unit, a de-echo unit, and a gain control unit, the feature extraction module includes an MFCC extraction unit and an LPCC extraction unit, the acoustic model module includes an HMM unit and a deep learning model unit, the language model module includes an N-gram model unit and a neural network language model unit, and the output module includes a text output unit and a speech synthesis unit.
[0008] Preferably, the output end of the sound sensor unit is electrically connected to the input end of the signal conversion unit, the output end of the noise reduction unit is electrically connected to the input end of the echo removal unit, the output end of the echo removal unit is electrically connected to the input end of the gain control unit, the output end of the MFCC extraction unit is electrically connected to the input end of the LPCC extraction unit, the output end of the HMM unit is electrically connected to the input end of the deep learning model unit, and the output end of the text output unit is electrically connected to the input end of the speech synthesis unit.
[0009] Preferably, the sound sensor unit uses a high-sensitivity microphone array to capture voice signals from different directions, and the signal conversion unit is used to convert the analog sound signal captured by the microphone array into a digital signal.
[0010] Preferably, the noise reduction unit adopts an adaptive filtering algorithm to suppress the noise in the audio signal, the echo removal unit adopts an echo cancellation algorithm to remove the echo component in the audio signal, the gain control unit is dynamically adjusted according to the strength of the audio signal to ensure that the signal maintains a constant amplitude during subsequent processing, the MFCC extraction unit adopts a Mel-frequency cepstral coefficient algorithm to extract MFCC features from the audio signal, and the LPCC extraction unit adopts a linear prediction cepstral coefficient algorithm to extract LPCC features from the audio signal.
[0011] Preferably, the HMM unit uses a hidden Markov model to model audio features and maps the audio features to phonemes or phoneme combinations, and the deep learning model unit uses a deep learning algorithm, such as a recurrent neural network, a long short-term memory network or a convolutional neural network, to model and identify audio features.
[0012] Preferably, the N-gram model unit uses an N-gram model to model vocabulary and grammar, and generates possible vocabulary sequences based on context information. The neural network language model unit uses a neural network algorithm to model vocabulary and semantics to generate more natural recognition results. The text output unit converts the recognized text sequence into character form, displays it on a screen or transmits it to other devices. The speech synthesis unit generates corresponding speech output based on the recognized text sequence.
[0013] Beneficial Effects
[0014] The present invention provides an AI-driven speech recognition system. Compared with the prior art, it has the following beneficial effects:
[0015] (1) By adopting advanced pre-processing technologies (such as noise reduction, echo removal and gain control) and feature extraction methods (such as MFCC and LPCC), the quality of audio signals can be significantly improved, thereby providing more reliable data input for subsequent acoustic models and language models. At the same time, the system also supports the training and optimization of acoustic models and language models, and continuously improves the accuracy and robustness of speech recognition through unsupervised learning, semi-supervised learning or supervised learning methods.
[0016] (2) The system is highly configurable and flexible. It can adjust parameters such as the adaptive filtering algorithm, echo cancellation algorithm, and automatic gain control algorithm, as well as select appropriate acoustic and language models according to different application scenarios and recognition requirements. This flexibility enables the system to be widely used in a variety of scenarios such as smart homes, in-car navigation, and smart customer service, meeting the voice interaction and recognition needs of different users.
[0017] (3) The output module converts the recognized text sequence into a readable form and supports speech synthesis, which can provide users with an intuitive and convenient voice interaction experience. At the same time, the system also has a feedback mechanism that can receive user feedback on the recognition results and adjust and optimize the system based on the feedback. This interactivity not only improves the user experience, but also helps the system to continuously learn and improve, providing users with more accurate and intelligent speech recognition services. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 is a system block diagram of the present invention;
[0019] Figure 2 is a schematic diagram of an audio input module of the present invention;
[0020] Figure 3 It is a schematic diagram of a pre-processing module of the present invention;
[0021] Figure 4 It is a schematic diagram of a feature extraction module of the present invention;
[0022] Figure 5 Schematic diagram of the acoustic model module of the present invention;
[0023] Figure 6 Schematic diagram of the language model module of the present invention;
[0024] Figure 7 It is a schematic diagram of the output module of the present invention.
[0025] In the figure: 01, audio input module; 02, preprocessing module; 03, feature extraction module; 04, acoustic model module; 05, language model module; 06, output module; 011, sound sensor unit; 012, signal conversion unit; 021, noise reduction unit; 022, echo removal unit; 023, gain control unit; 031, MFCC extraction unit; 032, LPCC extraction unit; 041, HMM unit; 042, deep learning model unit; 051, N-gram model unit; 052, neural network language model unit; 061, text output unit; 062, speech synthesis unit. DETAILED DESCRIPTION
[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0027] The present invention provides three technical solutions, including the following embodiments:
[0028] Example 1
[0029] See also Figure 1 , an AI-driven speech recognition system includes an audio input module 01, a preprocessing module 02, a feature extraction module 03, an acoustic model module 04, a language model module 05, and an output module 06. The audio input module 01 is used to capture and convert sound signals into digital signals, and is the starting point of the entire speech recognition system. The preprocessing module 02 cleans and formats the captured audio signals for subsequent processing. A voice activity detection (VAD) unit is added to accurately determine the starting and ending points of the speech signal and reduce the interference of non-speech segments on the recognition process. The feature extraction module 03 extracts features that are helpful for speech recognition from the preprocessed audio signal. The acoustic model module 04 is used to map audio features to phonemes or phoneme combinations, and is the core part of the speech recognition system. The language model module 05 is used to represent vocabulary, grammar, and semantic probability distributions, and helps the system generate more natural recognition results based on context information. The output module 06 converts the recognized text sequence into a readable form, such as characters, speech synthesis, etc.
[0030] The output end of the audio input module 01 is electrically connected to the input end of the preprocessing module 02, the output end of the preprocessing module 02 is electrically connected to the input end of the feature extraction module 03, the output end of the feature extraction module 03 is electrically connected to the input end of the acoustic model module 04, the output end of the acoustic model module 04 is electrically connected to the input end of the language model module 05, and the output end of the language model module 05 is electrically connected to the input end of the output module 06.
[0031] Example 2
[0032] Based on Example 1, see Figure 2-Figure 7 As shown, the audio input module 01 includes a sound sensor unit 011 and a signal conversion unit 012, the preprocessing module 02 includes a noise reduction unit 021, an echo removal unit 022, and a gain control unit 023, the feature extraction module 03 includes an MFCC extraction unit 031 and an LPCC extraction unit 032, the acoustic model module 04 includes an HMM unit 041 and a deep learning model unit 042, the language model module 05 includes an N-gram model unit 051 and a neural network language model unit 052, and the output module 06 includes a text output unit 061 and a speech synthesis unit 062.
[0033] The output end of the sound sensor unit 011 is electrically connected to the input end of the signal conversion unit 012, the output end of the noise reduction unit 021 is electrically connected to the input end of the echo removal unit 022, the output end of the echo removal unit 022 is electrically connected to the input end of the gain control unit 023, the output end of the MFCC extraction unit 031 is electrically connected to the input end of the LPCC extraction unit 032, the output end of the HMM unit 041 is electrically connected to the input end of the deep learning model unit 042, and the output end of the text output unit 061 is electrically connected to the input end of the speech synthesis unit 062.
[0034] Example 3
[0035] Based on Example 2, see Figure 2-Figure 7 As shown, the sound sensor unit 011 adopts a high-sensitivity microphone array to capture voice signals from different directions. The microphone array works together through multiple microphones and uses beamforming technology to improve the capture ability and directionality of voice signals. In addition to the high-sensitivity microphone array, noise suppression microphones and directional microphones can also be introduced to enhance the voice capture capability in noisy environments. The signal conversion unit 012 is used to convert the analog sound signal captured by the microphone array into a digital signal. The unit uses an analog-to-digital converter (ADC) for signal conversion to ensure that the converted digital signal has high fidelity and low noise.
[0036] The noise reduction unit 021 adopts an adaptive filtering algorithm to suppress the noise in the audio signal. The algorithm dynamically adjusts the parameters of the filter according to the characteristics and change rules of the noise, thereby effectively removing the background noise and improving the clarity of the speech signal. The echo removal unit 022 adopts an echo cancellation algorithm to remove the echo component in the audio signal. The algorithm effectively reduces the impact of the echo on speech recognition by estimating and eliminating the echo path. The gain control unit 023 dynamically adjusts according to the strength of the audio signal to ensure that the signal maintains a constant amplitude during subsequent processing. The unit adopts an automatic gain control (AGC) algorithm to automatically adjust the gain of the signal to avoid distortion caused by excessive or weak signals. The MFCC extraction unit 031 adopts a Mel frequency cepstral coefficient algorithm to extract MFCC features from the audio signal. MFCC is a feature parameter that can reflect the spectral characteristics of the speech signal and has an important influence on the accuracy of speech recognition. The LPCC extraction unit 032 adopts a linear prediction cepstral coefficient algorithm to extract LPCC features from the audio signal. LPCC is a feature parameter that can reflect the linear prediction characteristics of the speech signal and has an improvement effect on the robustness of speech recognition.
[0037] HMM unit 041 uses a hidden Markov model to model audio features and maps audio features to phonemes or phoneme combinations. HMM is a probability-based model that can capture the timing characteristics of speech signals and plays an important role in the accuracy of speech recognition. Deep learning model unit 042 uses deep learning algorithms, such as recurrent neural networks, long short-term memory networks or convolutional neural networks, to model and recognize audio features. Deep learning models have powerful nonlinear modeling capabilities and can capture the complex characteristics of speech signals, thereby improving the accuracy and robustness of speech recognition.
[0038] The N-gram model unit 051 uses the N-gram model to model vocabulary and grammar, and generates possible vocabulary sequences based on context information. The N-gram model is a statistical-based language model that can capture the correlation between vocabulary and improve the accuracy and fluency of speech recognition. The neural network language model unit 052 uses a neural network algorithm to model vocabulary and semantics to generate more natural recognition results. The neural network language model has powerful nonlinear modeling capabilities and can capture the complex relationship between vocabulary and improve the naturalness and accuracy of speech recognition. The text output unit 061 converts the recognized text sequence into character form, which is displayed on the screen or transmitted to other devices. The unit uses a standard text encoding format to ensure that the output text is readable and compatible. The speech synthesis unit 062 generates corresponding speech output based on the recognized text sequence. The unit uses a speech synthesis algorithm, such as text-to-speech (TTS) technology, to convert text information into speech signals to achieve a closed loop of speech interaction.
[0039] The system can be widely used in smart homes, car systems, telemedicine, online education and other fields, providing efficient and convenient voice interaction experience. For example, in smart homes, users can control home appliances through voice commands; in car systems, voice navigation, phone answering and other functions are realized; in telemedicine, doctors can input medical records through voice to improve work efficiency; in online education, students can ask questions through voice and get instant feedback.
[0040] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.
[0041] The above embodiments of the invention are described in detail, but the contents are only preferred embodiments of the invention and cannot be considered to limit the scope of the invention. All equivalent changes and improvements made within the scope of the invention should still fall within the scope of the invention.
Claims
1. An AI-driven speech recognition system, characterized by: The invention comprises an audio input module (01), a preprocessing module (02), a feature extraction module (03), an acoustic model module (04), a language model module (05), and an output module (06), wherein the audio input module (01) is used to capture and convert sound signals into digital signals, the preprocessing module (02) cleans and formats the captured audio signals, the feature extraction module (03) extracts features that are helpful for speech recognition from the preprocessed audio signals, the acoustic model module (04) is used to map audio features to phonemes or phoneme combinations, the language model module (05) is used to represent vocabulary, grammar and semantic probability distribution, and the output module (06) converts the recognized text sequence into a readable form.
2. The AI-driven speech recognition system according to claim 1, characterized in that: The output end of the audio input module (01) is electrically connected to the input end of the preprocessing module (02), the output end of the preprocessing module (02) is electrically connected to the input end of the feature extraction module (03), the output end of the feature extraction module (03) is electrically connected to the input end of the acoustic model module (04), the output end of the acoustic model module (04) is electrically connected to the input end of the language model module (05), and the output end of the language model module (05) is electrically connected to the input end of the output module (06).
3. The AI-driven speech recognition system according to claim 1, characterized in that: The audio input module (01) comprises a sound sensor unit (011) and a signal conversion unit (012); the preprocessing module (02) comprises a noise reduction unit (021), an echo removal unit (022), and a gain control unit (023); the feature extraction module (03) comprises an MFCC extraction unit (031) and an LPCC extraction unit (032); the acoustic model module (04) comprises an HMM unit (041) and a deep learning model unit (042); the language model module (05) comprises an N-gram model unit (051) and a neural network language model unit (052); and the output module (06) comprises a text output unit (061) and a speech synthesis unit (062).
4. The AI-driven speech recognition system according to claim 3, characterized in that: The output end of the sound sensor unit (011) is electrically connected to the input end of the signal conversion unit (012), the output end of the noise reduction unit (021) is electrically connected to the input end of the echo removal unit (022), the output end of the echo removal unit (022) is electrically connected to the input end of the gain control unit (023), the output end of the MFCC extraction unit (031) is electrically connected to the input end of the LPCC extraction unit (032), the output end of the HMM unit (041) is electrically connected to the input end of the deep learning model unit (042), and the output end of the text output unit (061) is electrically connected to the input end of the speech synthesis unit (062).
5. The AI-driven speech recognition system according to claim 3, characterized in that: The sound sensor unit (011) uses a high-sensitivity microphone array to capture voice signals from different directions, and the signal conversion unit (012) is used to convert the analog sound signal captured by the microphone array into a digital signal.
6. The AI-driven speech recognition system according to claim 3, characterized in that: The noise reduction unit (021) adopts an adaptive filtering algorithm to suppress noise in the audio signal, the echo removal unit (022) adopts an echo cancellation algorithm to remove echo components in the audio signal, the gain control unit (023) performs dynamic adjustment according to the strength of the audio signal to ensure that the signal maintains a constant amplitude during subsequent processing, the MFCC extraction unit (031) adopts a Mel-frequency cepstral coefficient algorithm to extract MFCC features from the audio signal, and the LPCC extraction unit (032) adopts a linear prediction cepstral coefficient algorithm to extract LPCC features from the audio signal.
7. The AI-driven speech recognition system according to claim 3, characterized in that: The HMM unit (041) uses a hidden Markov model to model audio features and maps the audio features to phonemes or phoneme combinations. The deep learning model unit (042) uses a deep learning algorithm, such as a recurrent neural network, a long short-term memory network or a convolutional neural network, to model and identify audio features.
8. The AI-driven speech recognition system according to claim 3, characterized in that: The N-gram model unit (051) uses an N-gram model to model vocabulary and grammar, and generates possible vocabulary sequences according to context information. The neural network language model unit (052) uses a neural network algorithm to model vocabulary and semantics, and generates more natural recognition results. The text output unit (061) converts the recognized text sequence into character form, which is displayed on a screen or transmitted to other devices. The speech synthesis unit (062) generates corresponding speech output according to the recognized text sequence.