Speech Recognition

In FunASR, we provide several ASR benchmarks, such as AISHLL, Librispeech, WenetSpeech, while different model architectures are supported, including conformer, paraformer, uniasr.

Quick Start

After downloaded and installed FunASR, users can use our provided recipes to easily reproduce the relevant experimental results. Here we take "paraformer on AISHELL-1" as an example.

First, move to the corresponding dictionary of the AISHELL-1 paraformer example.

cd egs/aishell/paraformer

Then you can directly start the recipe as follows:

conda activate funasr
. ./run.sh --CUDA_VISIBLE_DEVICES="0,1" --gpu_num=2

The training log files are saved in ${exp_dir}/exp/${model_dir}/log/train.log.*， which can be viewed using the following command:

vim exp/*_train_*/log/train.log.0

Users can observe the training loss, prediction accuracy and other training information, like follows:

... 1epoch:train:751-800batch:800num_updates: ... loss_ctc=106.703, loss_att=86.877, acc=0.029, loss_pre=1.552 ...
... 1epoch:train:801-850batch:850num_updates: ... loss_ctc=107.890, loss_att=87.832, acc=0.029, loss_pre=1.702 ...

At the end of each epoch, the evaluation metrics are calculated on the validation set, like follows:

... [valid] loss_ctc=99.914, cer_ctc=1.000, loss_att=80.512, acc=0.029, cer=0.971, wer=1.000, loss_pre=1.952, loss=88.285 ...

Also, users can use tensorboard to observe these training information by the following command:

tensorboard --logdir ${exp_dir}/exp/${model_dir}/tensorboard/train

Here is an example of loss:

The inference results are saved in ${exp_dir}/exp/${model_dir}/decode_asr_*/$dset. The main two files are text.cer and text.cer.txt. text.cer saves the comparison between the recognized text and the reference text, like follows:

...
BAC009S0764W0213(nwords=11,cor=11,ins=0,del=0,sub=0) corr=100.00%,cer=0.00%
ref:    构 建 良 好 的 旅 游 市 场 环 境
res:    构 建 良 好 的 旅 游 市 场 环 境
...

text.cer.txt saves the final results, like follows:

%WER ...
%SER ...
Scored ... sentences, ...

Introduction

We provide a recipe egs/aishell/paraformer/run.sh for training a paraformer model on AISHELL-1 dataset. This recipe consists of five stages, supporting training on multiple GPUs and decoding by CPU or GPU. Before introducing each stage in detail, we first explain several parameters which should be set by users.

CUDA_VISIBLE_DEVICES: 0,1 (Default), visible gpu list
gpu_num: 2 (Default), the number of GPUs used for training
gpu_inference: true (Default), whether to use GPUs for decoding
njob: 1 (Default),for CPU decoding, indicating the total number of CPU jobs; for GPU decoding, indicating the number of jobs on each GPU
raw_data: the raw path of AISHELL-1 dataset
feats_dir: the path for saving processed data
token_type: char (Default), indicate how to process text
type: sound (Default), set the input type
scp: wav.scp (Default), set the input file
nj: 64 (Default), the number of jobs for data preparation
speed_perturb: "0.9, 1.0 ,1.1" (Default), the range of speech perturbed
exp_dir: the path for saving experimental results
tag: exp1 (Default), the suffix of experimental result directory
stage 0 (Default), start the recipe from the specified stage
stop_stage 5 (Default), stop the recipe from the specified stage

Stage 0: Data preparation

This stage processes raw AISHELL-1 dataset $raw_data and generates the corresponding wav.scp and text in $feats_dir/data/xxx. xxx means train/dev/test. Here we assume users have already downloaded AISHELL-1 dataset. If not, users can download data here and set the path for $raw_data. The examples of wav.scp and text are as follows:

wav.scp

BAC009S0002W0122 /nfs/ASR_DATA/AISHELL-1/data_aishell/wav/train/S0002/BAC009S0002W0122.wav
BAC009S0002W0123 /nfs/ASR_DATA/AISHELL-1/data_aishell/wav/train/S0002/BAC009S0002W0123.wav
BAC009S0002W0124 /nfs/ASR_DATA/AISHELL-1/data_aishell/wav/train/S0002/BAC009S0002W0124.wav
...

text

BAC009S0002W0122 而 对 楼 市 成 交 抑 制 作 用 最 大 的 限 购
BAC009S0002W0123 也 成 为 地 方 政 府 的 眼 中 钉
BAC009S0002W0124 自 六 月 底 呼 和 浩 特 市 率 先 宣 布 取 消 限 购 后
...

These two files both have two columns, while the first column is wav ids and the second column is the corresponding wav paths/label tokens.

Stage 1: Feature and CMVN Generation

This stage computes CMVN based on train dataset, which is used in the following stages. Users can set nj to control the number of jobs for computing CMVN. The generated CMVN file is saved as $feats_dir/data/train/cmvn/am.mvn.

Stage 2: Dictionary Preparation

This stage processes the dictionary, which is used as a mapping between label characters and integer indices during ASR training. The processed dictionary file is saved as $feats_dir/data/$lang_toekn_list/$token_type/tokens.txt. An example of tokens.txt is as follows:

<blank>
<s>
</s>
一
丁
...
龚
龟
<unk>

There are four tokens must be specified:

<blank>: (required), indicates the blank token for CTC, must be in the first line
<s>: (required), indicates the start-of-sentence token, must be in the second line
</s>: (required), indicates the end-of-sentence token, must be in the third line
<unk>: (required), indicates the out-of-vocabulary token, must be in the last line

Stage 3: LM Training

Stage 4: ASR Training

This stage achieves the training of the specified model. To start training, users should manually set exp_dir to specify the path for saving experimental results. By default, the best $keep_nbest_models checkpoints on validation dataset will be averaged to generate a better model and adopted for decoding. FunASR implements train.py for training different models and users can configure the following parameters if necessary. The training command is as follows:

train.py \
    --task_name asr \
    --use_preprocessor true \
    --token_list $token_list \
    --data_dir ${feats_dir}/data \
    --train_set ${train_set} \
    --valid_set ${valid_set} \
    --data_file_names "wav.scp,text" \
    --cmvn_file ${feats_dir}/data/${train_set}/cmvn/am.mvn \
    --speed_perturb ${speed_perturb} \
    --resume true \
    --output_dir ${exp_dir}/exp/${model_dir} \
    --config $asr_config \
    --ngpu $gpu_num \
    ...

task_name: asr (Default), specify the task type of the current recipe
ngpu: 2 (Default), specify the number of GPUs for training. When ngpu > 1, DistributedDataParallel (DDP, the detail can be found here) training will be enabled. Correspondingly, CUDA_VISIBLE_DEVICES should be set to specify which ids of GPUs will be used.
use_preprocessor: true (Default), specify whether to use pre-processing on each sample
token_list: the path of token list for training
dataset_type: small (Default). FunASR supports small dataset type for training small datasets. Besides, an optional iterable-style DataLoader based on Pytorch Iterable-style DataPipes for large datasets is supported and users can specify dataset_type=large to enable it.
data_dir: the path of data. Specifically, the data for training is saved in $data_dir/data/$train_set while the data for validation is saved in $data_dir/data/$valid_set
data_file_names: "wav.scp,text" specify the speech and text file names for ASR
cmvn_file: the path of cmvn file
resume: true, whether to enable "checkpoint training"
output_dir: the path for saving training results
config: the path of configuration file, which is usually a YAML file in conf directory. In FunASR, the parameters of the training, including model, optimization, dataset, etc., can also be set in this file. Note that if the same parameters are specified in both recipe and config file, the parameters of recipe will be employed

Stage 5: Decoding

This stage generates the recognition results and calculates the CER to verify the performance of the trained model.

Mode Selection

As we support paraformer, uniasr, conformer and other models in FunASR, a mode parameter should be specified as asr/paraformer/uniasr according to the trained model.

Configuration

We support CTC decoding, attention decoding and hybrid CTC-attention decoding in FunASR, which can be specified by ctc_weight in a YAML file in conf directory. Specifically, ctc_weight=1.0 indicates CTC decoding, ctc_weight=0.0 indicates attention decoding, 0.0<ctc_weight<1.0 indicates hybrid CTC-attention decoding.

CPU/GPU Decoding

We support CPU and GPU decoding in FunASR. For CPU decoding, you should set gpu_inference=False and set njob to specify the total number of CPU decoding jobs. For GPU decoding, you should set gpu_inference=True. You should also set gpuid_list to indicate which GPUs are used for decoding and njobs to indicate the number of decoding jobs on each GPU.

Performance

We adopt CER to verify the performance. The results are in $exp_dir/exp/$model_dir/$decoding_yaml_name/$average_model_name/$dset, namely text.cer and text.cer.txt. text.cer saves the comparison between the recognized text and the reference text while text.cer.txt saves the final CER results. The following is an example of text.cer:

...
BAC009S0764W0213(nwords=11,cor=11,ins=0,del=0,sub=0) corr=100.00%,cer=0.00%
ref:    构 建 良 好 的 旅 游 市 场 环 境
res:    构 建 良 好 的 旅 游 市 场 环 境
...

Change settings

Here we explain how to perform common custom settings, which can help users to modify scripts according to their own needs.

Training with specified GPUs

For example, if users want to use 2 GPUs with id 2 and 3, users can run the following command:

. ./run.sh --CUDA_VISIBLE_DEVICES "2,3" --gpu_num 2

Start from/Stop at a specified stage

The recipe includes several stages. Users can start form or stop at any stage. For example, the following command achieves starting from the third stage and stopping at the fifth stage:

. ./run.sh --stage 3 --stop_stage 5

Specify total training steps

FunASR supports two parameters to specify the training steps, namely max_epoch and max_update. max_epoch indicates the total training epochs while max_update indicates the total training steps. If these two parameters are specified at the same time, once the training reaches any one of these two parameters, the training will be stopped.

Change the configuration of the model

The configuration of the model is set in the config file conf/train_*.yaml. Specifically, the default encoder configuration of paraformer is as follows:

encoder: conformer
encoder_conf:
    output_size: 256    # dimension of attention
    attention_heads: 4  # the number of heads in multi-head attention
    linear_units: 2048  # the number of units of position-wise feed forward
    num_blocks: 12      # the number of encoder blocks
    dropout_rate: 0.1
    positional_dropout_rate: 0.1
    attention_dropout_rate: 0.0
    input_layer: conv2d  # encoder input layer architecture type
    normalize_before: true
    pos_enc_layer_type: rel_pos
    selfattention_layer_type: rel_selfattn
    activation_type: swish
    macaron_style: true
    use_cnn_module: true
    cnn_module_kernel: 15

Users can change the encoder configuration by modify these values. For example, if users want to use an encoder with 16 conformer blocks and each block has 8 attention heads, users just need to change num_blocks from 12 to 16 and change attention_heads from 4 to 8. Besides, the batch_size, learning rate and other training hyper-parameters are also set in this config file. To change these hyper-parameters, users just need to directly change the corresponding values in this file. For example, the default learning rate is 0.0005. If users want to change the learning rate to 0.0002, set the value of lr as lr: 0.0002.

Change different input data type

FunASR supports different input data types, including sound, kaldi_ark, npy, text and text_int. Users can specify any number and any type of input, which is achieved by data_names and data_types (in config/train_*.yaml). For example, ASR task usually requires speech and the transcripts as input. In FunASR, by default, speech is saved as raw audio (such as wav format) and transcripts are saved as text format. Correspondingly, data_names and data_types are set as follows (seen in config/train_*.yaml):

dataset_conf:
    data_names: speech,text
    data_types: sound,text
    ...

When the input type changes to FBank, users just need to modify as data_types: kaldi_ark,text in the config file. Note data_file_names used in train.py should also be changed to the new file name.

How to resume training process

FunASR supports resuming training as follows:

train.py ... --resume true ...

How to transfer / fine-tuning from pre-trained models

FunASR supports transferring / fine-tuning from a pre-trained model by specifying the init_param parameter. The usage format is as follows:

train.py ... --init_param <file_path>:<src_key>:<dst_key>:<exclude_keys>  ..

For example, the following command achieves loading all pretrained parameters starting from decoder except decoder.embed and set it to model.decoder2:

train.py ... --init_param model.pb:decoder:decoder2:decoder.embed  ...

Besides, loading parameters from multiple pre-trained models is supported. For example, the following command achieves loading encoder parameters from the pre-trained model1 and decoder parameters from the pre-trained model2:

train.py ... --init_param model1.pb:encoder  --init_param model2.pb:decoder ...

How to freeze part of the model parameters

In certain situations, users may want to fix part of the model parameters update the rest model parameters. FunASR employs freeze_param to achieve this. For example, to fix all parameters like encoder.*, users need to set freeze_param as follows:

train.py ... --freeze_param encoder ...

ModelScope Usage

Users can use ModelScope for inference and fine-tuning based on a trained academic model. To achieve this, users need to run the stage 6 in the script. In this stage, relevant files required by ModelScope will be generated automatically. Users can then use the corresponding ModelScope interface by replacing the model name with the local trained model path. For the detailed usage of the ModelScope interface, please refer to ModelScope Usage.

Decoding by CPU or GPU

We support CPU and GPU decoding. For CPU decoding, set gpu_inference=false and njob to specific the total number of CPU jobs. For GPU decoding, first set gpu_inference=true. Then set gpuid_list to specific which GPUs for decoding and njob to specific the number of decoding jobs on each GPU.

asr_recipe.md 15 KB Histórico Raw