| | |
| | | ([简体中文](./README_zh.md)|English) |
| | | |
| | | # Speech Recognition |
| | | |
| | | > **Note**: |
| | | > The modelscope pipeline supports all the models in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/modelscope_models.html#pretrained-models-on-modelscope) to inference and finetine. Here we take the typic models as examples to demonstrate the usage. |
| | | > **Note**: |
| | | > The modelscope pipeline supports all the models in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/model_zoo/modelscope_models.html#pretrained-models-on-modelscope) to inference and finetine. Here we take the typic models as examples to demonstrate the usage. |
| | | |
| | | ## Inference |
| | | |
| | |
| | | rec_result = inference_pipeline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav') |
| | | print(rec_result) |
| | | ``` |
| | | #### [Paraformer-online Model](https://www.modelscope.cn/models/damo/speech_paraformer_asr_nat-zh-cn-16k-common-vocab8404-online/summary) |
| | | #### [Paraformer-online Model](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online/summary) |
| | | ##### Streaming Decoding |
| | | ```python |
| | | inference_pipeline = pipeline( |
| | | task=Tasks.auto_speech_recognition, |
| | | model='damo/speech_paraformer_asr_nat-zh-cn-16k-common-vocab8404-online', |
| | | model='damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online', |
| | | model_revision='v1.0.7', |
| | | update_model=False, |
| | | mode='paraformer_streaming' |
| | | ) |
| | | import soundfile |
| | | speech, sample_rate = soundfile.read("example/asr_example.wav") |
| | | |
| | | param_dict = {"cache": dict(), "is_final": False} |
| | | chunk_stride = 7680# 480ms |
| | | # first chunk, 480ms |
| | | speech_chunk = speech[0:chunk_stride] |
| | | chunk_size = [0, 10, 5] #[0, 10, 5] 600ms, [0, 8, 4] 480ms |
| | | encoder_chunk_look_back = 4 #number of chunks to lookback for encoder self-attention |
| | | decoder_chunk_look_back = 1 #number of encoder chunks to lookback for decoder cross-attention |
| | | param_dict = {"cache": dict(), "is_final": False, "chunk_size": chunk_size, |
| | | "encoder_chunk_look_back": encoder_chunk_look_back, "decoder_chunk_look_back": decoder_chunk_look_back} |
| | | chunk_stride = chunk_size[1] * 960 # 600ms、480ms |
| | | # first chunk, 600ms |
| | | speech_chunk = speech[0:chunk_stride] |
| | | rec_result = inference_pipeline(audio_in=speech_chunk, param_dict=param_dict) |
| | | print(rec_result) |
| | | # next chunk, 480ms |
| | | # next chunk, 600ms |
| | | speech_chunk = speech[chunk_stride:chunk_stride+chunk_stride] |
| | | rec_result = inference_pipeline(audio_in=speech_chunk, param_dict=param_dict) |
| | | print(rec_result) |
| | | ``` |
| | | |
| | | ##### Fake Streaming Decoding |
| | | ```python |
| | | from modelscope.pipelines import pipeline |
| | | from modelscope.utils.constant import Tasks |
| | | |
| | | inference_pipeline = pipeline( |
| | | task=Tasks.auto_speech_recognition, |
| | | model='damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online', |
| | | model_revision='v1.0.7', |
| | | update_model=False, |
| | | mode="paraformer_fake_streaming" |
| | | ) |
| | | audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav' |
| | | rec_result = inference_pipeline(audio_in=audio_in) |
| | | print(rec_result) |
| | | ``` |
| | | Full code of demo, please ref to [demo](https://github.com/alibaba-damo-academy/FunASR/discussions/241) |
| | | |
| | | #### [Paraformer-contextual Model](https://www.modelscope.cn/models/damo/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404/summary) |
| | | ```python |
| | | from modelscope.pipelines import pipeline |
| | | from modelscope.utils.constant import Tasks |
| | | |
| | | param_dict = dict() |
| | | # param_dict['hotword'] = "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/hotword.txt" |
| | | param_dict['hotword']="邓郁松 王颖春 王晔君" |
| | | inference_pipeline = pipeline( |
| | | task=Tasks.auto_speech_recognition, |
| | | model="damo/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404", |
| | | param_dict=param_dict) |
| | | |
| | | rec_result = inference_pipeline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_hotword.wav') |
| | | print(rec_result) |
| | | ``` |
| | | |
| | | #### [UniASR Model](https://www.modelscope.cn/models/damo/speech_UniASR_asr_2pass-zh-cn-8k-common-vocab3445-pytorch-online/summary) |
| | | There are three decoding mode for UniASR model(`fast`、`normal`、`offline`), for more model detailes, please refer to [docs](https://www.modelscope.cn/models/damo/speech_UniASR_asr_2pass-zh-cn-8k-common-vocab3445-pytorch-online/summary) |
| | | There are three decoding mode for UniASR model(`fast`、`normal`、`offline`), for more model details, please refer to [docs](https://www.modelscope.cn/models/damo/speech_UniASR_asr_2pass-zh-cn-8k-common-vocab3445-pytorch-online/summary) |
| | | ```python |
| | | decoding_model = "fast" # "fast"、"normal"、"offline" |
| | | inference_pipeline = pipeline( |
| | |
| | | ``` |
| | | The decoding mode of `fast` and `normal` is fake streaming, which could be used for evaluating of recognition accuracy. |
| | | Full code of demo, please ref to [demo](https://github.com/alibaba-damo-academy/FunASR/discussions/151) |
| | | |
| | | #### [Paraformer-Spk](https://modelscope.cn/models/damo/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn/summary) |
| | | This model allows user to get recognition results which contain speaker info of each sentence. Refer to [CAM++](https://modelscope.cn/models/damo/speech_campplus_speaker-diarization_common/summary) for detailed information about speaker diarization model. |
| | | ```python |
| | | from modelscope.pipelines import pipeline |
| | | from modelscope.utils.constant import Tasks |
| | | |
| | | if __name__ == '__main__': |
| | | audio_in = 'https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_speaker_demo.wav' |
| | | output_dir = "./results" |
| | | inference_pipeline = pipeline( |
| | | task=Tasks.auto_speech_recognition, |
| | | model='damo/speech_paraformer-large-vad-punc-spk_asr_nat-zh-cn', |
| | | model_revision='v0.0.2', |
| | | vad_model='damo/speech_fsmn_vad_zh-cn-16k-common-pytorch', |
| | | punc_model='damo/punc_ct-transformer_cn-en-common-vocab471067-large', |
| | | output_dir=output_dir, |
| | | ) |
| | | rec_result = inference_pipeline(audio_in=audio_in, batch_size_token=5000, batch_size_token_threshold_s=40, max_single_segment_time=6000) |
| | | print(rec_result) |
| | | ``` |
| | | |
| | | #### [RNN-T-online model]() |
| | | Undo |
| | | |
| | | #### API-reference |
| | | ##### Define pipeline |
| | | #### [MFCCA Model](https://www.modelscope.cn/models/NPU-ASLP/speech_mfcca_asr-zh-cn-16k-alimeeting-vocab4950/summary) |
| | | For more model details, please refer to [docs](https://www.modelscope.cn/models/NPU-ASLP/speech_mfcca_asr-zh-cn-16k-alimeeting-vocab4950/summary) |
| | | ```python |
| | | from modelscope.pipelines import pipeline |
| | | from modelscope.utils.constant import Tasks |
| | | |
| | | inference_pipeline = pipeline( |
| | | task=Tasks.auto_speech_recognition, |
| | | model='NPU-ASLP/speech_mfcca_asr-zh-cn-16k-alimeeting-vocab4950', |
| | | model_revision='v3.0.0' |
| | | ) |
| | | |
| | | rec_result = inference_pipeline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav') |
| | | print(rec_result) |
| | | ``` |
| | | |
| | | ### API-reference |
| | | #### Define pipeline |
| | | - `task`: `Tasks.auto_speech_recognition` |
| | | - `model`: model name in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/modelscope_models.html#pretrained-models-on-modelscope), or model path in local disk |
| | | - `model`: model name in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/model_zoo/modelscope_models.html#pretrained-models-on-modelscope), or model path in local disk |
| | | - `ngpu`: `1` (Default), decoding on GPU. If ngpu=0, decoding on CPU |
| | | - `ncpu`: `1` (Default), sets the number of threads used for intraop parallelism on CPU |
| | | - `ncpu`: `1` (Default), sets the number of threads used for intraop parallelism on CPU |
| | | - `output_dir`: `None` (Default), the output path of results if set |
| | | - `batch_size`: `1` (Default), batch size when decoding |
| | | ##### Infer pipeline |
| | | - `audio_in`: the input to decode, which could be: |
| | | #### Infer pipeline |
| | | - `audio_in`: the input to decode, which could be: |
| | | - wav_path, `e.g.`: asr_example.wav, |
| | | - pcm_path, `e.g.`: asr_example.pcm, |
| | | - pcm_path, `e.g.`: asr_example.pcm, |
| | | - audio bytes stream, `e.g.`: bytes data from a microphone |
| | | - audio sample point,`e.g.`: `audio, rate = soundfile.read("asr_example_zh.wav")`, the dtype is numpy.ndarray or torch.Tensor |
| | | - wav.scp, kaldi style wav list (`wav_id \t wav_path``), `e.g.`: |
| | | - wav.scp, kaldi style wav list (`wav_id \t wav_path`), `e.g.`: |
| | | ```text |
| | | asr_example1 ./audios/asr_example1.wav |
| | | asr_example2 ./audios/asr_example2.wav |
| | |
| | | ### Inference with multi-thread CPUs or multi GPUs |
| | | FunASR also offer recipes [egs_modelscope/asr/TEMPLATE/infer.sh](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/asr/TEMPLATE/infer.sh) to decode with multi-thread CPUs, or multi GPUs. |
| | | |
| | | - Setting parameters in `infer.sh` |
| | | - `model`: model name in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/modelscope_models.html#pretrained-models-on-modelscope), or model path in local disk |
| | | - `data_dir`: the dataset dir needs to include `wav.scp`. If `${data_dir}/text` is also exists, CER will be computed |
| | | - `output_dir`: output dir of the recognition results |
| | | - `batch_size`: `64` (Default), batch size of inference on gpu |
| | | - `gpu_inference`: `true` (Default), whether to perform gpu decoding, set false for CPU inference |
| | | - `gpuid_list`: `0,1` (Default), which gpu_ids are used to infer |
| | | - `njob`: only used for CPU inference (`gpu_inference`=`false`), `64` (Default), the number of jobs for CPU decoding |
| | | - `checkpoint_dir`: only used for infer finetuned models, the path dir of finetuned models |
| | | - `checkpoint_name`: only used for infer finetuned models, `valid.cer_ctc.ave.pb` (Default), which checkpoint is used to infer |
| | | #### Settings of `infer.sh` |
| | | - `model`: model name in [model zoo](https://alibaba-damo-academy.github.io/FunASR/en/model_zoo/modelscope_models.html#pretrained-models-on-modelscope), or model path in local disk |
| | | - `data_dir`: the dataset dir needs to include `wav.scp`. If `${data_dir}/text` is also exists, CER will be computed |
| | | - `output_dir`: output dir of the recognition results |
| | | - `batch_size`: `64` (Default), batch size of inference on gpu |
| | | - `gpu_inference`: `true` (Default), whether to perform gpu decoding, set false for CPU inference |
| | | - `gpuid_list`: `0,1` (Default), which gpu_ids are used to infer |
| | | - `njob`: only used for CPU inference (`gpu_inference`=`false`), `64` (Default), the number of jobs for CPU decoding |
| | | - `checkpoint_dir`: only used for infer finetuned models, the path dir of finetuned models |
| | | - `checkpoint_name`: only used for infer finetuned models, `valid.cer_ctc.ave.pb` (Default), which checkpoint is used to infer |
| | | - `decoding_mode`: `normal` (Default), decoding mode for UniASR model(fast、normal、offline) |
| | | - `hotword_txt`: `None` (Default), hotword file for contextual paraformer model(the hotword file name ends with .txt") |
| | | |
| | | - Decode with multi GPUs: |
| | | #### Decode with multi GPUs: |
| | | ```shell |
| | | bash infer.sh \ |
| | | --model "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch" \ |
| | |
| | | --gpu_inference true \ |
| | | --gpuid_list "0,1" |
| | | ``` |
| | | - Decode with multi-thread CPUs: |
| | | #### Decode with multi-thread CPUs: |
| | | ```shell |
| | | bash infer.sh \ |
| | | --model "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch" \ |
| | |
| | | --njob 64 |
| | | ``` |
| | | |
| | | - Results |
| | | #### Results |
| | | |
| | | The decoding results can be found in `$output_dir/1best_recog/text.cer`, which includes recognition results of each sample and the CER metric of the whole test set. |
| | | |
| | |
| | | [finetune.py](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/asr/TEMPLATE/finetune.py) |
| | | ```python |
| | | import os |
| | | |
| | | from modelscope.metainfo import Trainers |
| | | from modelscope.trainers import build_trainer |
| | | from modelscope.msdatasets.audio.asr_dataset import ASRDataset |
| | | |
| | | from funasr.datasets.ms_dataset import MsDataset |
| | | from funasr.utils.modelscope_param import modelscope_args |
| | | |
| | | |
| | | def modelscope_finetune(params): |
| | | if not os.path.exists(params.output_dir): |
| | | os.makedirs(params.output_dir, exist_ok=True) |
| | | # dataset split ["train", "validation"] |
| | | ds_dict = ASRDataset.load(params.data_path, namespace='speech_asr') |
| | | ds_dict = MsDataset.load(params.data_path) |
| | | kwargs = dict( |
| | | model=params.model, |
| | | data_dir=ds_dict, |
| | |
| | | work_dir=params.output_dir, |
| | | batch_bins=params.batch_bins, |
| | | max_epoch=params.max_epoch, |
| | | lr=params.lr) |
| | | lr=params.lr, |
| | | mate_params=params.param_dict) |
| | | trainer = build_trainer(Trainers.speech_asr_trainer, default_args=kwargs) |
| | | trainer.train() |
| | | |
| | | |
| | | if __name__ == '__main__': |
| | | from funasr.utils.modelscope_param import modelscope_args |
| | | params = modelscope_args(model="damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch") |
| | | params.output_dir = "./checkpoint" # 模型保存路径 |
| | | params.data_path = "speech_asr_aishell1_trainsets" # 数据路径,可以为modelscope中已上传数据,也可以是本地数据 |
| | | params.dataset_type = "small" # 小数据量设置small,若数据量大于1000小时,请使用large |
| | | params.batch_bins = 2000 # batch size,如果dataset_type="small",batch_bins单位为fbank特征帧数,如果dataset_type="large",batch_bins单位为毫秒, |
| | | params.max_epoch = 50 # 最大训练轮数 |
| | | params.lr = 0.00005 # 设置学习率 |
| | | |
| | | params = modelscope_args(model="damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch", data_path="./data") |
| | | params.output_dir = "./checkpoint" # m模型保存路径 |
| | | params.data_path = "speech_asr_aishell1_trainsets" # 数据路径 |
| | | params.dataset_type = "small" # 小数据量设置small,若数据量大于1000小时,请使用large |
| | | params.batch_bins = 2000 # batch size,如果dataset_type="small",batch_bins单位为fbank特征帧数,如果dataset_type="large",batch_bins单位为毫秒, |
| | | params.max_epoch = 20 # 最大训练轮数 |
| | | params.lr = 0.00005 # 设置学习率 |
| | | init_param = [] # 初始模型路径,默认加载modelscope模型初始化,例如: ["checkpoint/20epoch.pb"] |
| | | freeze_param = [] # 模型参数freeze, 例如: ["encoder"] |
| | | ignore_init_mismatch = True # 是否忽略模型参数初始化不匹配 |
| | | use_lora = False # 是否使用lora进行模型微调 |
| | | params.param_dict = {"init_param":init_param, "freeze_param": freeze_param, "ignore_init_mismatch": ignore_init_mismatch} |
| | | if use_lora: |
| | | enable_lora = True |
| | | lora_bias = "all" |
| | | lora_params = {"lora_list":['q','v'], "lora_rank":8, "lora_alpha":16, "lora_dropout":0.1} |
| | | lora_config = {"enable_lora": enable_lora, "lora_bias": lora_bias, "lora_params": lora_params} |
| | | params.param_dict.update(lora_config) |
| | | |
| | | modelscope_finetune(params) |
| | | ``` |
| | | |
| | |
| | | - `batch_bins`: batch size. For dataset_type is `small`, `batch_bins` indicates the feature frames. For dataset_type is `large`, `batch_bins` indicates the duration in ms |
| | | - `max_epoch`: number of training epoch |
| | | - `lr`: learning rate |
| | | - `init_param`: `[]`(Default), init model path, load modelscope model initialization by default. For example: ["checkpoint/20epoch.pb"] |
| | | - `freeze_param`: `[]`(Default), Freeze model parameters. For example:["encoder"] |
| | | - `ignore_init_mismatch`: `True`(Default), Ignore size mismatch when loading pre-trained model |
| | | - `use_lora`: `False`(Default), Fine-tuning model use lora, more detail please refer to [LORA](https://arxiv.org/pdf/2106.09685.pdf) |
| | | |
| | | - Training data formats: |
| | | ```sh |