System and method for transforming audio using a foundational generative machine learning model
A foundational generative machine learning model addresses the inefficiencies of multiple specialized models by using multi-modal prompts to perform various audio transformation tasks within a single framework, enhancing efficiency and flexibility in audio processing.
Patent Information
- Application Number
- PCT/CN2023/109536
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-07-27
- Publication Date
- 2025-01-30
AI Technical Summary
Existing audio transformation technologies require multiple specialized machine learning models to perform specific audio processing tasks, making them inefficient and inflexible in meeting diverse customer needs.
A foundational generative machine learning model that uses multi-modal prompts combining audio and text tokens to perform a range of audio and voice generation and transformation tasks, including text-to-speech, noise reduction, personalization, and voice conversion, within a single model framework.
The solution enables a single model to effectively handle diverse audio transformation needs, including tasks that require both audio and text inputs, thereby improving efficiency and flexibility in audio processing applications.
Smart Images

Figure CN2023109536_30012025_PF_FP_ABST
Abstract
Description
System and Method for Transforming Audio Using a Foundational Generative Machine Learning ModelBackground
[0001] The field of generative artificial intelligence (AI) has been experiencing a tremendous surge in the text and vision domains, bringing about a significant transformation across various technological fields. Inspired by these advancements, audio and speech-generative models are being explored with a focus on a multitude of methodologies using large scale data. For example, many specialized machine learning models are generally used to perform specific audio processing tasks (e.g., one model for cloning human voices; another model for text-to-speech processing; etc. ) . Therefore, a number of separate machine learning models are required to audio transformation services that serve customers with various needs.Brief Description of the Drawings
[0002] FIG. 1 is a flow chart of one implementation of an audio transformation process;
[0003] FIG. 2 is a diagrammatic view of the audio transformation process of FIG. 1 during training of a generative machine learning model;
[0004] FIG. 3 is a flow chart of one implementation of an audio transformation process during training of a generative machine learning model;
[0005] FIGS. 4A-5B are diagrammatic views of various task-dependent prompts generated according to example implementations of the audio generation and transformation process of FIG. 1;
[0006] FIG. 6 is a diagrammatic view of the audio generation or transformation process of FIG. 1 during inference; and
[0007] FIG. 7 is a diagrammatic view of computer system and an audio generation and transformation process coupled to a distributed computing network.
[0008] Like reference symbols in the various drawings indicate like elements.
[0009] Detailed Description of the Preferred Embodiments
[0010] Implementations of the present disclosure provide a foundational model dedicated for voice and sound generation and transformations. The foundational model realizes a range of audio and voice generation and transformation effects, using multi-modal prompts that incorporate audio and text tokens where the tokens combine to define the specifics of the output audio to be generated. By accepting these multi-modal prompts, the foundational model is able to effectively meet diverse audio generation and transformation needs, including text-to-speech, noise reduction / control, personalization, voice conversion / overdubbing / morphing, and, more importantly, their combined effects. The foundational model’s capabilities extend to recordings obtained in various acoustic conditions, allowing it to offer a comprehensive solution for a wide range of voice and sound transformation requirements.
[0011] For example, implementations of the present disclosure allow for a single foundational generative machine learning model (e.g., a multi-modal generative machine learning model) to perform various transformation tasks on audio data using a combination of an autoregressive machine learning model and a non-autoregressive machine learning model that generate and process task-dependent prompts. Specifically, the process converts audio data into a plurality of neural codec tokens and generates a task-dependent prompt for the multi-modal generative machine learning model by encoding a token-based representation of a task to be performed on the audio data into the plurality of neural codec tokens. Generating the task-dependent prompt for the multi-modal generative machine learning model includes combining the token-based representation of the task into the plurality of neural codec tokens in a predefined sequence based upon, at least in part, the task. The multi-modal generative machine learning model also processes multi-modal inputs (e.g., audio and text) to perform particular tasks (e.g., TTS) . For example, the multi-modal generative machine learning model generates a semantic prompt by converting input text data into a plurality of semantic tokens and is trained for a particular task involving by processing the task- dependent prompt, the semantic prompt, and the plurality of neural codec tokens. In this manner, a single foundational generative machine learning model is able to perform various types of audio generation and transformation tasks.
[0012] The details of one or more implementations are set forth in the accompanying drawings and the description below. Other features and advantages will become apparent from the description, the drawings, and the claims.
[0013] The Audio Transformation Process:
[0014] Referring to FIGS. 1-6, audio transformation process 10 determines 100 a task to be performed on the audio data using a generative machine learning model. A token-based representation of the task to be performed on the audio data is generated 102. Audio data is converted 104 into a plurality of neural codec tokens. A task-dependent prompt for a generative machine learning model is generated 106 by encoding a token-based representation of a task to be performed on the audio data into the plurality of neural codec tokens. Audio transformation process 10 causes 108 the task-dependent prompt to be processed using the generative machine learning model to accomplish the specified task.
[0015] Referring also to FIG. 2, audio transformation process 10 allows for the training and inference-based use of a foundational multi-modal generative machine learning model for various tasks based upon, at least in part, a task-dependent prompt provided to the multi-modal generative machine learning model. A generative machine learning model (e.g., generative machine learning model 200) is configured to receive input prompts and / or example entries and / or contextual information concerning a request to generate an audio response. Generative machine learning model 200 includes a neural network with many parameters (typically millions or billions of weights or more) , trained on large quantities of unlabeled and / or labeled data using self-supervised learning, semi-supervised learning, and / or fine-tuning of the weights to cater the neural network for particular tasks or workloads.
[0016] In some implementations, generative machine learning model 200 includes an autoregressive neural codec language model and a non-autoregressive neural codec language model. In one example, generative machine learning model 200 is built upon the VALL-E-based TTS model which demonstrates effective zero-shot generation capabilities. The VALL-E model can generate personalized and high-quality speech output based on a text input (i.e., a semantic prompt) and a short recording of a target speaker’s voice (i.e., an acoustic prompt) where the target speaker may be unseen during training. Generative machine learning model 200 may include two sub-modules using neural codecs based on residual vector quantization: an autoregressive (AR) neural codec language model (e.g., autoregressive neural codec language model 202) for predicting the tokens of the first codebook of the neural codec with the acoustic and semantic prompts, and a non-autoregressive neural codec language model (e.g., non-autoregressive neural codec language model 204) for producing the tokens of the remaining codebooks of the neural codec.
[0017] Multi-modal Generative Machine Learning Model Training
[0018] Referring also to FIG. 3, audio transformation process 10 determines 100 a task to be performed on audio data using a generative machine learning model. Audio transformation process 10 may use tasks (e.g., task 206) to define how generative machine learning model 200 is trained and used to transform audio data. In some implementations, task 206 includes one or more of: noise suppression; noise estimation; target speaker extraction; quiet speech editing; noisy speech editing; and text-to-speech processing. During training, a task (e.g., task 206) is defined beforehand or randomly selected and provided to audio transformation process 10 for each training sample or a group of samples.
[0019] In another example, during inference, a task can be specified explicitly. For example, audio transformation process 10 provides a user interface for a user to select a type of task for generative machine learning model 200 to execute. In another example, audio transformation process 10 receives a description of a task from a machine learning model or other process and audio transformation process 10 converts that description into task 206. For instance, a user may use a chat interface of a large language model (LLM) with a text or speech prompt to perform a particular task. In this example, audio transformation process 10 receives task 206 from the LLM or converts a prompt from the LLM to define task 206.
[0020] In some implementations, audio transformation process 10 generates 102 a token-based representation of the task to be performed on the audio data. For example, audio transformation process 10 converts the task (e.g., task 206) into a task token (including any number of task tokens) . In one example, audio transformation process 10 uses a task token generator (e.g., task token generator 208) to convert task 206 into a token-based representation of task 206 (e.g., task token 210) . In some implementations, task token generator is a machine learning model configured or trained to convert a description or other representation of a task (e.g., task 206) into task token 210. In one example, task token 210 may be a code or other numerical representation of a task. In one example, task token 210 is generated from task 206 as a numerical code or identifier describing task 206.
[0021] In some implementations, audio transformation process 10 obtains 300 the audio data for the generative machine learning model based upon, at least in part, the task to be performed. For example, when processing task 206, audio transformation process 10 identifies audio data to use for training generative machine learning model 200. In one example, audio transformation process 10 provides task token 210 to a task-based neural codec extractor (e.g., task-based neural codec extractor 212) . Task-based neural codec extractor 212 is a hardware and / or software component that processes task token 210 to determine which audio data to obtain for generative machine learning model 200. For example, different tasks may involve different combinations of training data (e.g., training data 214) . In one example, audio transformation process 10 determines 100 task 206 to be performed and generates 102 task token 210. Using task token 210, audio transformation process 10 uses task-based neural codec extractor 212 to obtain 300 the audio data for generative machine learning model 200. For example, using task token 210, audio transformation process 10 determines the types of training data (e.g., training data 214) needed for task 206. Training data 214 includes any combination of databases or accessible data sources for types of audio data to be used to train generative machine learning model 200. In one example, training data 214 includes databases of corresponding clean and noisy data (e.g., clean data 216 and noisy data 218) . Each portion of clean data 216 has a corresponding noisy portion in noisy data 218. In another example, training data 214 includes noise data 220 which can be used to train generative machine learning model 200. In response to processing task token 210 using task-based neural codec extractor 212, audio transformation process 10 obtains 300 audio data (e.g., audio data 222) for training generative machine learning model 200.
[0022] In another example, audio transformation process 10 obtains 300 audio data to process with task 206 along with task 206. For example, when audio transformation process 10 obtains task 206, audio transformation process 10 also obtains 300 audio data to process with task 206. In this manner, audio transformation process 10 provides task 206 and the received or input audio data (e.g., audio data 222) to task-based neural codex extractor 212. Received audio data 222 is processed by task-based neural codec extractor 212. As will be discussed in greater detail below, audio transformation process 10 converts audio data 222 into a plurality of neural codec tokens for processing by generative machine learning model 200.
[0023] In some implementations, audio transformation process 10 converts 104 audio data into a plurality of neural codec tokens. In one example, the audio data (e.g., audio data 222) includes speech data (i.e., audio data with at least one speech component) . In another example, audio data 222 includes non-speech data (i.e., audio data without any speech components) . A neural codec token is a discrete audio code representative of a portion or sample of audio data 222. During processing, audio transformation process 10 converts 104 audio data 222 into a plurality of neural code tokens for processing by generative machine learning model 200 and converts output neural codec tokens produced by generative machine learning model 200 back into a waveform representation.
[0024] In some implementations, converting 104 the audio data into the plurality of neural codec tokens includes processing 110 the audio data with a neural codec token extractor. A neural codec token extractor (e.g., neural codec extractor 212) is a machine learning model that encodes waveforms from audio data 222 into discrete acoustic codes (e.g., plurality of neural codec tokens 224) . In one example, the neural codec token extractor is the EnCodec model. In some implementations, neural codec extractor 208 is an encoder-decoder architecture based on residual vector quantization using “L” quantization layers (where “L” is a number) . In one example, there are eight quantization layers and each layer of neural codec extractor 208 produces discrete audio codes consisting of e.g., 1024 entries at a sampling rate of e.g., 75 Hz.
[0025] In some implementations, audio transformation process 10 generates 106 a task-dependent prompt for a generative machine learning model by encoding a token-based representation of a task to be performed on the audio data into the plurality of neural codec tokens. A task-dependent prompt is a sequence of tokens that describe the task to be performed on the token and the audio data to process. In some implementations, audio transformation process 10 provides task token 210 and plurality of neural codec tokens 224 to a task-dependent prompt generator (e.g., task-dependent prompt generator 226) . Task-dependent prompt generator 226 is a hardware and / or software component configured to generate 106 the task-dependent prompt (e.g., task dependent prompt 228) by encoding the token-based representation of the task (e.g., task token 210) into plurality of neural codec tokens 224.
[0026] In some implementations, encoding the token-based representation of the task (e.g., task token 210) includes combining 112 the token-based representation of the task into the plurality of neural codec tokens in a predefined sequence based upon, at least in part, the task. For example, audio transformation process 10 combines 112 the plurality of neural codec tokens (e.g., neural codec token 224) and task token 210 in a particular sequence depending upon task 206 to generate task-dependent prompt. For example, different task-dependent prompts are generated 106 for a particular task by combining 112 task token 210 and plurality of neural codec tokens 224. Examples of task-dependent prompts include prompts for noise suppression, noise estimation, target speech extraction, quiet speech editing, noisy speech editing, and zero-shot TTS. However, it will be appreciated that these are provided only as examples and various task-dependent prompts may be generated for other prompts within the scope of the present disclosure.
[0027] Referring also to FIG. 4A and in one example, suppose task 206 concerns noise suppression, and audio transformation process 10 converts audio data 222 (which includes clean audio data and noisy audio data) into neural codec tokens 400, 402, 404, 406 representing noisy audio data, neural codec token 408 representing clean audio data, and task 206 into task token 210. While FIG. 4 shows four neural codec tokens, it will be appreciated that any number of neural codec tokens may be generated within the scope of the present disclosure. Audio transformation process 10 combines 112 task token 210 into plurality of neural codec tokens 400, 402, 404, 406. For the noise suppression task (e.g., task 206) , audio transformation process 10 adds task token 210 to neural codec tokens 400, 402, 404, 406 representing the noisy audio data to generate 106 task-dependent prompt (e.g., task-dependent prompt 410) . Additionally, for training generative machine learning model 200, audio transformation process 10 also converts clean audio data into neural codec tokens 408. Using the combination of task-dependent prompt 410 describing the task of noise suppression and including noisy audio data and neural codec tokens 408 for the clean audio data, generative machine learning model 200 can be trained to perform noise suppression by outputting a sequence of neural codec tokens 408 of the clean audio data from input neural codec tokens 400, 402, 404, 406 of the noisy audio data.
[0028] Referring also to FIG. 4B and in another example, suppose task 206 concerns noise estimation and audio transformation process 10 converts audio data 222 (which includes noisy audio data) into neural codec tokens 400, 402, 404, 406 representing noisy audio data, neural codec token 412 representing noise audio data, and task 206 into task token 210. Audio transformation process 10 combines 112 task token 210 into plurality of neural codec tokens 400, 402, 404, 406. For the noise estimation task (e.g., task 206) , audio transformation process 10 adds task token 210 to neural codec tokens 400, 402, 404, 406 representing the noisy audio data to generate 106 task-dependent prompt (e.g., task-dependent prompt 414) . Additionally, for training generative machine learning model 200, audio transformation process 10 also converts noise audio data into neural codec tokens 412. Using the combination of task-dependent prompt 412 describing the task of noise estimation and including noisy audio data and neural codec tokens 412 for the noise audio data to be estimated, generative machine learning model 200 can be trained to perform noise estimation by outputting a sequence of neural codec tokens 412 of the noise audio data from input neural codec tokens 400, 402, 404, 406. In some implementations, a noise estimation task is without text guidance (e.g., there is not input text data corresponding to the audio data) . In this example and as described above, the noise estimation task is performed on generative machine learning model 200 without including the “guidance” of input text data to estimate noise in the audio data. In some implementations, the noise estimation task is with text guidance (e.g., there is input text data corresponding to the audio data) . In this example and as described above, the noise estimation task is performed on generative machine learning model 200 by including the “guidance” of input text data to estimate noise in the audio data.
[0029] Referring also to FIG. 4C and in another example, suppose task 206 concerns target speaker extraction and audio transformation process 10 converts audio data 222 into neural codec tokens 400, 402 representing speaker enrollment audio data and neural codec tokens 404, 406 representing multi-speaker mixed audio to be enhanced, neural codec token 408 representing clean audio data, and task 206 into task token 210. Audio transformation process 10 combines 112 task token 210 into plurality of neural codec tokens 400, 402, 404, 406. For the target speaker extraction task (e.g., task 206) , audio transformation process 10 combines or adds task token 210 after neural codec tokens 400, 402 representing the speaker enrollment audio data but before neural codec tokens 404, 406 representing the multi-speaker mixed audio to be enhanced to generate 106 task-dependent prompt (e.g., task-dependent prompt 416) . Using the combination of task-dependent prompt 416 (describing the task of target speaker extraction including speaker enrollment audio data and multi-speaker mixed audio data to be enhanced) and neural codec tokens 408 for the clean audio data, generative machine learning model 200 can be trained to perform target speaker extraction by outputting a sequence of neural codec tokens 408 of the clean audio data from input neural codec tokens 400, 402 representing speaker enrollment audio data and 404, 406 representing multi-speaker audio data to be enhanced. In some implementations, a target speaker extraction task is without text guidance (e.g., there is not input text data corresponding to the audio data) . In this example and as described above, the target speaker extraction task is performed on generative machine learning model 200 without including the “guidance” of input text data to extract target speaker audio from the audio data. In some implementations, the target speaker extraction task is with text guidance (e.g., there is input text data corresponding to the audio data) . In this example and as described above, the target speaker extraction task is performed on generative machine learning model 200 by including the “guidance” of input text data to extract target speaker audio from the audio data.
[0030] As described above, noise suppression, and target speaker extraction are representative audio enhancement tasks where the clean speech signal to be extracted may be partially or entirely overlapped with interfering signals, such as background noise or other speakers’ voices. Unlike generative tasks, the aim is to recover the original underlying clean signal with high fidelity. Consequently, the task-dependent prompt contains tokens representing the whole audio to be enhanced for generative machine learning model 200 to accomplish the specified task, whereas a semantic prompt is optional. To accommodate the scenarios where the text of the target audio is unavailable, audio transformation process 10 may omit a part or the entirety of the semantic prompt when processing the noise suppression, and target speaker extraction training batches during training. In one example, audio transformation process 10 omits the semantic prompt at a 50%probability so that generative machine learning model 200 equally experiences both text and text-less scenarios. As will be discussed in greater detail below, a semantic prompt is obtained by applying Grapheme-to-Phoneme (G2P) or other phoneme conversion schemes to a transcription of the audio data, where the transcription for the audio data may be provided by automated speech recognition (ASR) or human annotators.
[0031] Regarding speech editing, audio transformation process 10 further categorizes the task into two scenarios: quiet editing and noisy editing. These categories are distinguished based on whether the audio to be edited contains a threshold amount of background noise or not. Referring also to FIG. 4D and in another example, suppose task 206 concerns quiet speech editing (e.g., editing of speech audio data with less than a predefined threshold of noise) , and audio transformation process 10 converts audio data 222 into neural codec token 400 representing a portion of audio data to be edited and neural codec tokens 402, 404 representing portions of the audio data that are not to be edited, neural codec token 408 representing clean audio data, and task 206 into task tokens 418, 420. Audio transformation process 10 combines 112 task tokens 418, 420 into plurality of neural codec tokens 400, 402, 404. For the quiet editing task (e.g., task 206) , audio transformation process 10 combines or adds task token 418 representing the start of editing in the sequence of neural codec tokens before a mask token that replaces neural codec token 400 (e.g., mask token 420) , and token 422 representing the end of editing after mask token 420 and before neural codec tokens 402, 404 to generate 106 task-dependent prompt (e.g., task-dependent prompt 424) . Using the combination of task-dependent prompt 424 (describing the task of quiet speech editing including a token representing the start of editing in the sequence of neural codec tokens, a mask token in place of a neural codec token to be edited, and a token representing the end of editing) and neural codec tokens 408 for the clean audio data, generative machine learning model 200 can be trained to perform quiet speech editing by outputting a sequence of neural codec tokens 408 from input neural codec tokens 400, 402, 404.
[0032] In the noisy editing case, audio transformation process 10 keeps the background noise unaltered while changing the speech content based on a text input. Consequently, audio transformation process 10 does not use a mask token as it hides the background noise information to be preserved. Referring also to FIG. 5A and in another example, suppose task 206 concerns noisy speech editing (e.g., editing of speech audio data with more than a predefined threshold of noise) and audio transformation process 10 converts audio data 222 into neural codec token 400 representing a portion of audio data to be edited and neural codec tokens 402, 404 representing portions of the audio data that are not to be edited, neural codec token 408 representing clean audio data, and task 206 into task tokens 500, 502. Audio transformation process 10 combines 112 task tokens 500, 502 into plurality of neural codec tokens 400, 402, 404. For the noisy editing task (e.g., task 206) , audio transformation process 10 combines or adds task token 500 representing the start of editing in the sequence of neural codec tokens before the neural codec token to be edited (e.g., neural codec token 400) , and token 502 representing the end of editing after neural codec token 400 and before neural codec tokens 402, 404 to generate 106 task-dependent prompt (e.g., task-dependent prompt 504) . Using the combination of task-dependent prompt 504 (describing the task of quiet speech editing including a token representing the start of editing in the sequence of neural codec tokens and a token representing the end of editing) and neural codec tokens 408 for the clean audio data, generative machine learning model 200 can be trained to perform noisy speech editing by outputting a sequence of neural codec tokens 408 from input neural codec tokens 400, 402, 404.
[0033] For the task of zero-shot text-to-speech (TTS) , audio transformation process 10 includes both a semantic prompt and acoustic tokens (e.g., neural codec tokens) . Zero-shot text-to-speech is the application of zero-shot learning to text-to-speech processing. Zero-shot learning generally works by associating observed and non-observed classes through a form of auxiliary information, which encodes observable distinguishing properties of objects. In text-to-speech, zero-shot learning is applied for generating “unseen” speakers. Referring again to FIG. 2 and in some implementations, audio transformation process 10 generates 114 a semantic prompt by converting input text data into a plurality of semantic tokens. For example, audio transformation process 10 process input text data (e.g., input text data 230) using semantic prompt generator 232 to generate a plurality of semantic tokens (e.g., plurality of semantic tokens 234) . In one example, input text data 230 is obtained by applying ASR to audio data 222 to be generated. In another example, input text data 230 is from a transcription of audio data 222. In some implementations, semantic prompt generator 232 is a hardware and / or software component configured to convert input text data 232 into a plurality of phonemes and from the plurality of phonemes to a plurality of semantic tokens (e.g., plurality of semantic tokens 234) . In one example, semantic prompt generator 232 uses a Grapheme-to-Phoneme (G2P) conversion method. In some implementations, audio transformation process 10 provides plurality of semantic tokens 234 as prompt (e.g., semantic prompt) to generative machine learning model 200.
[0034] Referring also to FIG. 5B and in one example, suppose task 206 concerns zero-shot TTS and audio transformation process 10 converts audio data 222 (which includes clean audio data) into neural codec tokens 506, 508, 510, 512 representing limited samples of clean audio data (i.e., enrollment data) , neural codec token 408 representing clean audio data, task 206 into task token 210, and input text data 230 into semantic prompt 234. Audio transformation process 10 uses task token 210 to perform text-to-speech processing using input text data 234 and plurality of neural codec tokens 506, 508, 510, 512. For the zero-shot TTS task (e.g., task 206) , audio transformation process 10 does not add or combine task token 210 before or after neural codec tokens 506, 508, 510, 512 representing the clean audio data samples (e.g., three-second samples) to generate 106 task-dependent prompt (e.g., task-dependent prompt 514) . Additionally, for training generative machine learning model 200, audio transformation process 10 also converts clean audio data into neural codec tokens 408. Using the combination of task-dependent prompt 514 describing the task of zero-shot TTS, semantic prompt 234, and neural codec tokens 408 for the clean audio data, generative machine learning model 200 can be trained to perform zero-shot TTS by outputting a sequence of neural codec tokens 408 from input text data 234.
[0035] In some implementations, audio transformation process 10 trains 302 the generative machine learning model by processing the task-dependent prompt using the generative machine learning model. For example, during training, each generative machine learning model update iteration uses a mini batch (e.g., batch of corresponding audio data types) obtained from a randomly chosen data configuration and the data configuration changes over the iterations. Each mini batch is tailored specifically for a particular task, effectively training the generative machine learning model on different aspects or objectives. This approach helps generative machine learning model 200 to learn and adapt to a diverse range of tasks simultaneously. Within each mini-batch, task-dependent prompt generator 226 is used to create task-dependent prompt 228 that is fed into generative machine learning model 200. Task-dependent prompt generator 226 is designed to be adaptable to the specific requirements of each task by inserting task tokens, ensuring that generative machine learning model 200 receives appropriate and relevant input for accomplishing the given task.
[0036] In some implementations, audio transformation process 10 trains 304 the generative machine learning model by processing the task-dependent prompt and the semantic prompt using the generative machine learning model. For example, generative machine learning model 200 makes use of both AR and NAR machine learning models. Specifically, AR machine learning model 202 is trained using the token sequence obtained from the first quantization layer of neural codec extractor 208. The AR machine learning model 202 comprises a stack of transformer decoder layers and is optimized by maximizing the log-likelihood of the generated output. It is important to note that the AR machine learning model 202 in generative machine learning model 200 is conditioned on the concatenated embeddings of both the acoustic and semantic prompts. NAR machine learning model 204 starts with the AR machine learning model 202’s output and proceeds to iteratively generate the remaining tokens for the other quantization layers of neural codec extractor 208. Both components employ a decoder-only structure and utilize in-context learning with two prompts: plurality of semantic tokens 234 (e.g., semantic prompt) and task-dependent prompt 228. In some implementations, after processing task-dependent prompt 228 and plurality of semantic tokens 234, generative machine learning model 200 generates a plurality of output neural codec tokens (e.g., plurality of output neural codec tokens 236) .
[0037] Multi-modal Generative Machine Learning Model Inference
[0038] Referring also to FIG. 6, during inference, audio transformation process 10 uses generative machine learning model 200 as a foundational model to undertake various types of tasks for a particular audio input and / or text input. For example, task tokens (e.g., task token 210) are inserted into task-dependent prompt 228, enabling generative machine learning model 200 to automatically perform a specified task (e.g., task 206) accordingly. In some implementations, audio transformation process 10 is able to translate task 206 from an input prompt received from a user. For example, a user may provide a prompt including task 206, audio data 222, and input text data 230. In one example, audio transformation process 10 receives a selection of task 206, audio data 222, and input text data 230 using a user interface. In another example, audio transformation process 10 receives a selection of task 206, audio data 222, and input text data 230 using chat-based machine learning model (e.g., a chat-based LLM) . As discussed above, audio transformation process 10 converts 104 audio data 222 into plurality of neural codec tokens 224.
[0039] In some implementations, audio transformation process 10 generates 106 task-dependent prompt 228 for generative machine learning model 200 by encoding a token-based representation of task 206 to be performed on audio data 222 into plurality of neural codec tokens 224. Specifically, audio transformation process 10 combines 112 plurality of neural codec tokens 224 with task token 210 to generate 106 task-dependent prompt 228. As discussed above, the sequence of how task tokens 210 are combined with or added to neural codec tokens 224 is based upon, at least in part, task 206.
[0040] In some implementations, audio transformation process 10 causes 108 the processing of the task-dependent prompt using the generative machine learning model. For example, audio transformation process 10 provides task-dependent prompt 228 and plurality of semantic tokens 234 for processing by generative machine learning model 200. As discussed above, audio transformation process 10 uses the combination of AR machine learning model 202 and NAR machine learning model 204 to generate a plurality of output neural codec tokens (e.g., plurality of output neural codec tokens 236) . In some implementations, audio transformation process 10 converts plurality of output neural codec tokens 236 by processing plurality of output neural coded tokens 236 with a waveform generator (e.g., waveform generator 600) . Waveform generator 600 is a hardware and / or software component that converts neural codec tokens into audio data samples. In some implementations, audio transformation process 10 converts each of plurality of output neural codec tokens 236 into separate audio data samples. Audio transformation process 10 combines or aggregates the separate audio data samples to form audio data 602. As discussed above, audio data 602 represents the output of generative machine learning model 200 after processing audio data 222 for a defined task (e.g., task 206) . In this manner, generative machine learning model 200 is able to transform audio data according to various tasks without requiring multiple machine learning models.
[0041] In a comparison of generative machine learning model 200 and specific targeted machine learning models for performing particular tasks (e.g., DCCRN, VoiceFilter, and VALL-E) , the following results were observed comparing word error rate (WER) and deep noise suppression mean opinion score (DNSMOS) as shown below in Tables 1-3 for different tasks:
[0042] Table 1: Noise Suppression
[0043] Table 2: Target Speaker Extraction
[0044] Table 3: Zero-shot TTS
[0045] As shown above, generative machine learning model 200 as trained using audio transformation process 10, performed comparably to specific targeted machine learning models for various tasks (e.g., noise suppression, target speaker extraction, and zero-shot TTS) . In this manner, audio transformation process 10 allows for generative machine learning model 200 to process different tasks for audio generation or transformation with similar performance metrics (e.g., WER and DNSMOS) as machine learning models trained exclusively for specific tasks.
[0046] System Overview:
[0047] Referring to FIG. 7, an audio transformation process 10 is shown to reside on and is executed by storage system 700, which is connected to network 702 (e.g., the Internet or a local area network) . Examples of storage system 700 include: a Network Attached Storage (NAS) system, a Storage Area Network (SAN) , a personal computer with a memory system, a server computer with a memory system, and a cloud-based device with a memory system. A SAN includes one or more of a personal computer, a server computer, a series of server computers, a minicomputer, a mainframe computer, a RAID device, and a NAS system.
[0048] The various components of storage system 700 execute one or more operating systems, examples of which include: OS Red Mobile, Chrome OS, Blackberry OS, Fire OS, or a custom operating system (Microsoft and Windows are registered trademarks of Microsoft Corporation in the United States, other countries or both; Mac and OS X are registered trademarks of Apple Inc. in the United States, other countries or both; Red Hat is a registered trademark of Red Hat Corporation in the United States, other countries or both; and Linux is a registered trademark of Linus Torvalds in the United States, other countries or both) .
[0049] The instruction sets and subroutines of audio transformation process 10, which are stored on storage device 704 included within storage system 700, are executed by one or more processors (not shown) and one or more memory architectures (not shown) included within storage system 700. Storage device 704 may include: a hard disk drive; an optical drive; a RAID device; a random-access memory (RAM) ; a read-only memory (ROM) ; and all forms of flash memory storage devices. Additionally or alternatively, some portions of the instruction sets and subroutines of audio transformation process 10 are stored on storage devices (and / or executed by processors and memory architectures) that are external to storage system 700.
[0050] In some implementations, network 702 is connected to one or more secondary networks (e.g., network 706) , examples of which include: a local area network; a wide area network; or an intranet.
[0051] Various input / output (IO) requests (e.g., IO request 708) are sent from client applications 710, 712, 714, 716 to storage system 700. Examples of IO request 708 include data write requests (e.g., a request that content be written to storage system 700) and data read requests (e.g., a request that content be read from storage system 700) .
[0052] The instruction sets and subroutines of client applications 710, 712, 714, 716, which may be stored on storage devices 718, 720, 722, 724 (respectively) coupled to client electronic devices 726, 728, 730, 732 (respectively) , may be executed by one or more processors (not shown) and one or more memory architectures (not shown) incorporated into client electronic devices 726, 728, 730, 732 (respectively) . Storage devices 718, 720, 722, 724 may include: hard disk drives; tape drives; optical drives; RAID devices; random access memories (RAM) ; read-only memories (ROM) , and all forms of flash memory storage devices. Examples of client electronic devices 726, 728, 730, 732 include personal computer 726, laptop computer 728, smartphone 730, laptop computer 732, a server (not shown) , a data-enabled, and a dedicated network device (not shown) . Client electronic devices 726, 728, 730, 732 each execute an operating system.
[0053] Users 734, 736, 738, 740 may access storage system 700 directly through network 702 or through secondary network 706. Further, storage system 700 may be connected to network 702 through secondary network 706, as illustrated with link line 742.
[0054] The various client electronic devices may be directly or indirectly coupled to network 702 (or network 706) . For example, personal computer 726 is shown directly coupled to network 702 via a hardwired network connection. Further, laptop computer 732 is shown directly coupled to network 706 via a hardwired network connection. Laptop computer 728 is shown wirelessly coupled to network 702 via wireless communication channel 744 established between laptop computer 728 and wireless access point (e.g., WAP) 746, which is shown directly coupled to network 702. WAP 746 may be, for example, an IEEE 802.11a, 802.11b, 802.11g, 802.11n, and / or device that is capable of establishing a wireless communication channel 744 between laptop computer 728 and WAP 746. Smartphone 730 is shown wirelessly coupled to network 702 via wireless communication channel 748 established between smartphone 730 and cellular network / bridge 750, which is shown directly coupled to network 702.
[0055] General:
[0056] As will be appreciated by one skilled in the art, the present disclosure may be embodied as a method, a system, or a computer program product. Accordingly, the present disclosure may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, micro-code, etc. ) or an embodiment combining software and hardware aspects that may all generally be referred to herein as a “circuit, ” “module” or “system. ” Furthermore, the present disclosure may take the form of a computer program product on a computer-usable storage medium having computer-usable program code embodied in the medium.
[0057] Any suitable computer usable or computer readable medium may be used. The computer-usable or computer-readable medium may be, for example but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, device, or propagation medium. More specific examples (anon- exhaustive list) of the computer-readable medium may include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM) , a read-only memory (ROM) , an erasable programmable read-only memory (EPROM or Flash memory) , an optical fiber, a portable compact disc read-only memory (CD-ROM) , an optical storage device, a transmission media such as those supporting the Internet or an intranet, or a magnetic storage device. The computer-usable or computer-readable medium may also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, via, for instance, optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner, if necessary, and then stored in a computer memory. In the context of this document, a computer-usable or computer-readable medium may be any medium that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-usable medium may include a propagated data signal with the computer-usable program code embodied therewith, either in baseband or as part of a carrier wave. The computer usable program code may be transmitted using any appropriate medium, including but not limited to the Internet, wireline, optical fiber cable, RF, etc.
[0058] Computer program code for carrying out operations of the present disclosure may be written in an object-oriented programming language. However, the computer program code for carrying out operations of the present disclosure may also be written in conventional procedural programming languages, such as the “C” programming language or similar programming languages. The program code may execute entirely on the user’s computer, partly on the user’s computer, as a stand-alone software package, partly on the user’s computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user’s computer through a local area network / a wide area network / the Internet.
[0059] The present disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer / special purpose computer / other programmable data processing apparatus, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.
[0060] These computer program instructions may also be stored in a computer-readable memory that may direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function / act specified in the flowchart and / or block diagram block or blocks.
[0061] The computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.
[0062] The flowcharts and block diagrams in the figures may illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function (s) . It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, not at all, or in any combination with any other flowcharts depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations of blocks in the block diagrams and / or flowchart illustrations, may be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
[0063] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. As used herein, the singular forms “a” , “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” and / or “comprising, ” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.
[0064] The corresponding structures, materials, acts, and equivalents of all means or step plus function elements in the claims below are intended to include any structure, material, or act for performing the function in combination with other claimed elements as specifically claimed. The description of the present disclosure has been presented for purposes of illustration and description but is not intended to be exhaustive or limited to the disclosure in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the disclosure. The embodiment was chosen and described in order to best explain the principles of the disclosure and the practical application, and to enable others of ordinary skill in the art to understand the disclosure for various embodiments with various modifications as are suited to the particular use contemplated.
[0065] A number of implementations have been described. Having thus described the disclosure of the present application in detail and by reference to embodiments thereof, it will be apparent that modifications and variations are possible without departing from the scope of the disclosure defined in the appended claims.
Claims
1.A computer-implemented method, executed on a computing device, comprising:determining a task to be performed on audio data using a generative machine learning model;generating a token-based representation of the task to be performed on the audio data;converting the audio data into a plurality of neural codec tokens;generating a task-dependent prompt for the generative machine learning model by encoding the token-based representation of the task to be performed on the audio data into the plurality of neural codec tokens; andcausing processing of the task-dependent prompt by the generative machine learning model.2.The computer-implemented method of claim 1, wherein the audio data includes speech data.3.The computer-implemented method of claim 1, wherein the generative machine learning model includes an autoregressive neural codec language model and a non-autoregressive neural codec language model.4.The computer-implemented method of claim 1, wherein converting the audio data into the plurality of neural codec tokens includes processing the audio data with a neural codec token extractor.5.The computer-implemented method of claim 1, wherein generating the task-dependent prompt for the generative machine learning model includes combining the token-based representation of the task into the plurality of neural codec tokens in a predefined sequence based upon, at least in part, the task.6.The computer-implemented method of claim 1, wherein the task includes one or more of:noise suppression;noise estimation;target speaker extraction;quiet speech editing;noisy speech editing; andtext-to-speech processing.7.The computer-implemented method of claim 1, further comprising:generating a semantic prompt by converting input text data into a plurality of semantic tokens.8.A computing system comprising:a memory; anda processor configured to determine a task to be performed on audio data using a generative machine learning model, to obtain the audio data for the generative machine learning model based upon, at least in part, the task to be performed, to convert the audio data into a plurality of neural codec tokens, to generate a task-dependent prompt for the generative machine learning model by encoding a token-based representation of the task to be performed into the plurality of neural codec tokens, and to train the generative machine learning model by processing the task-dependent prompt using the generative machine learning model.9.The computing system of claim 8, wherein the audio data includes speech data.10.The computing system of claim 8, wherein the generative machine learning model includes an autoregressive neural codec language model and a non-autoregressive neural codec language model.11.The computing system of claim 8, wherein converting the audio data into the plurality of neural codec tokens includes processing the audio data with a neural codec token extractor.12.The computing system of claim 8, wherein generating the task-dependent prompt for the generative machine learning model includes combining the token-based representation of the task into the plurality of neural codec tokens in a predefined sequence based upon, at least in part, the task.13.The computing system of claim 8, wherein the task includes one or more of:noise suppression;noise estimation without text guidance;noise estimation with text guidance;target speaker extraction without text guidance;target speaker extraction with text guidance;quiet speech editing;noisy speech editing; andtext-to-speech processing.14.The computing system of claim 8, wherein the processor is further configured to:generate a semantic prompt by converting input text data into a plurality of semantic tokens; andtrain the generative machine learning model by processing the task-dependent prompt, the semantic prompt, and the plurality of neural codec tokens using the generative machine learning model.15.A computer program product residing on a computer readable medium having a plurality of instructions stored thereon which, when executed by a processor, cause the processor to perform operations comprising:determining a task to be performed on the audio data using a multi-modal generative machine learning model;generating a token-based representation of the task to be performed on the audio data;converting the audio data into a plurality of neural codec tokens;generating a task-dependent prompt for the multi-modal generative machine learning model by encoding the token-based representation of the task to be performed on the audio data into the plurality of neural codec tokens;generate a semantic prompt by converting text data into a plurality of semantic tokens; andproviding the task-dependent prompt and the semantic prompt for processing by the multi-modal generative machine learning model.
Citation Information
Cited By
Speech enhancement method and device, training method and device, computer equipment and storage medium
CN121034331A
Encoder training method, audio generation method and audio retrieval method
CN121789697A
Electronic device and audio signal link design method
GB2702302A