Merge pull request #3242 from jiamingkong/develop

Adding WavLM implementation

README.md

@@ -178,6 +178,7 @@ Via the easy-to-use, efficient, flexible and scalable implementation, our vision
   - 🧩  *Cascaded models application*: as an extension of the typical traditional audio tasks, we combine the workflows of the aforementioned tasks with other fields like Natural language processing (NLP) and Computer Vision (CV).
 
 ### Recent Update
+- 👑 2023.05.31: Add [WavLM ASR-en](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/examples/librispeech/asr5), WavLM fine-tuning for ASR on LibriSpeech.
 - 👑 2023.05.04: Add [HuBERT ASR-en](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/examples/librispeech/asr4), HuBERT fine-tuning for ASR on LibriSpeech.
 - ⚡ 2023.04.28: Fix [0-d tensor](https://github.com/PaddlePaddle/PaddleSpeech/pull/3214), with the upgrade of paddlepaddle==2.5, the problem of modifying 0-d tensor has been solved.
 - 👑 2023.04.25: Add [AMP for U2 conformer](https://github.com/PaddlePaddle/PaddleSpeech/pull/3167).

README_cn.md

@@ -183,6 +183,8 @@
   - 🧩 级联模型应用: 作为传统语音任务的扩展，我们结合了自然语言处理、计算机视觉等任务，实现更接近实际需求的产业级应用。
 
 ### 近期更新
+- 👑 2023.05.31: 新增 [WavLM ASR-en](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/examples/librispeech/asr5), 基于WavLM的英语识别微调，使用LibriSpeech数据集
+- 👑 2023.05.04: 新增 [HuBERT ASR-en](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/examples/librispeech/asr4), 基于HuBERT的英语识别微调，使用LibriSpeech数据集
 - ⚡ 2023.04.28: 修正 [0-d tensor](https://github.com/PaddlePaddle/PaddleSpeech/pull/3214), 配合PaddlePaddle2.5升级修改了0-d tensor的问题。
 - 👑 2023.04.25: 新增 [U2 conformer 的 AMP 训练](https://github.com/PaddlePaddle/PaddleSpeech/pull/3167).
 - 👑 2023.04.06: 新增 [srt格式字幕生成功能](./demos/streaming_asr_server)。

demos/speech_ssl/README.md

@@ -36,7 +36,7 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/en.wav
   ```
   Arguments:
   - `input`(required): Audio file to recognize.
-  - `model`: Model type of asr task. Default: `wav2vec2`, choices: [wav2vec2, hubert].
+  - `model`: Model type of asr task. Default: `wav2vec2`, choices: [wav2vec2, hubert, wavlm].
   - `task`: Output type. Default: `asr`.
   - `lang`: Model language. Default: `en`.
   - `sample_rate`: Sample rate of the model. Default: `16000`.

demos/speech_ssl/README_cn.md

@@ -36,7 +36,7 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/en.wav
   ```
   参数：
   - `input`(必须输入)：用于识别的音频文件。
-  - `model`：ASR 任务的模型，默认值：`wav2vec2`, 可选项：[wav2vec2, hubert]。
+  - `model`：ASR 任务的模型，默认值：`wav2vec2`, 可选项：[wav2vec2, hubert, wavlm]。
   - `task`：输出类别，默认值：`asr`。
   - `lang`：模型语言，默认值：`en`。
   - `sample_rate`：音频采样率，默认值：`16000`。

examples/librispeech/asr5/README.md

@@ -0,0 +1,197 @@
+# WavLM2ASR with Librispeech
+This example contains code used to finetune [WavLM](https://arxiv.org/abs/2110.13900) model with [Librispeech dataset](http://www.openslr.org/resources/12)
+## Overview
+All the scripts you need are in `run.sh`. There are several stages in `run.sh`, and each stage has its function.
+| Stage | Function                                                     |
+|:---- |:----------------------------------------------------------- |
+| 0     | Process data. It includes: <br>       (1) Download the dataset <br>       (2) Calculate the CMVN of the train dataset <br>       (3) Get the vocabulary file <br>       (4) Get the manifest files of the train, development and test dataset<br>       (5) Download the pretrained wav2vec2 model |
+| 1     | Train the model                                              |
+| 2     | Get the final model by averaging the top-k models, set k = 1 means to choose the best model |
+| 3     | Test the final model performance                             |
+| 4     | Infer the single audio file                                  |
+
+
+You can choose to run a range of stages by setting `stage` and `stop_stage `. 
+
+For example, if you want to execute the code in stage 2 and stage 3, you can run this script:
+```bash
+bash run.sh --stage 2 --stop_stage 3
+```
+Or you can set `stage` equal to `stop-stage` to only run one stage.
+For example, if you only want to run `stage 0`, you can use the script below:
+```bash
+bash run.sh --stage 0 --stop_stage 0
+```
+The document below will describe the scripts in `run.sh` in detail.
+## The Environment Variables
+The path.sh contains the environment variables. 
+```bash
+. ./path.sh
+. ./cmd.sh
+```
+This script needs to be run first. And another script is also needed:
+```bash
+source ${MAIN_ROOT}/utils/parse_options.sh
+```
+It will support the way of using `--variable value` in the shell scripts.
+## The Local Variables
+Some local variables are set in `run.sh`. 
+`gpus` denotes the GPU number you want to use. If you set `gpus=`, it means you only use CPU. 
+`stage` denotes the number of stages you want to start from in the experiments.
+`stop stage` denotes the number of the stage you want to end at in the experiments. 
+`conf_path` denotes the config path of the model.
+`avg_num` denotes the number K of top-K models you want to average to get the final model.
+`audio file` denotes the file path of the single file you want to infer in stage 5
+`ckpt` denotes the checkpoint prefix of the model, e.g. "WavLMASR"
+
+You can set the local variables (except `ckpt`) when you use `run.sh`
+
+For example, you can set the `gpus` and `avg_num` when you use the command line:
+```bash
+bash run.sh --gpus 0,1 --avg_num 20
+```
+## Stage 0: Data Processing
+To use this example, you need to process data firstly and you can use stage 0 in `run.sh` to do this. The code is shown below:
+```bash
+ if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+     # prepare data
+     bash ./local/data.sh || exit -1
+ fi
+```
+Stage 0 is for processing the data.
+
+If you only want to process the data. You can run
+```bash
+bash run.sh --stage 0 --stop_stage 0
+```
+You can also just run these scripts in your command line.
+```bash
+. ./path.sh
+. ./cmd.sh
+bash ./local/data.sh
+```
+After processing the data, the `data` directory will look like this:
+```bash
+data/
+|-- dev.meta
+|-- lang_char
+|   `-- bpe_unigram_5000.model
+|   `-- bpe_unigram_5000.vocab
+|   `-- vocab.txt
+|-- manifest.dev
+|-- manifest.dev.raw
+|-- manifest.test
+|-- manifest.test.raw
+|-- manifest.train
+|-- manifest.train.raw
+|-- mean_std.json
+|-- test.meta
+`-- train.meta
+```
+
+Stage 0 also downloads the pre-trained [wavlm](https://paddlespeech.bj.bcebos.com/wavlm/wavlm-base-plus.pdparams) model.
+```bash
+mkdir -p exp/wavlm
+wget -P exp/wavlm https://paddlespeech.bj.bcebos.com/wavlm/wavlm-base-plus.pdparams
+```
+## Stage 1: Model Training
+If you want to train the model. you can use stage 1 in `run.sh`. The code is shown below. 
+```bash
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+     # train model, all `ckpt` under `exp` dir
+     CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${ckpt}
+ fi
+```
+If you want to train the model, you can use the script below to execute stage 0 and stage 1:
+```bash
+bash run.sh --stage 0 --stop_stage 1
+```
+or you can run these scripts in the command line (only use CPU).
+```bash
+. ./path.sh
+. ./cmd.sh
+bash ./local/data.sh
+CUDA_VISIBLE_DEVICES= ./local/train.sh conf/wavlmASR.yaml wavlmASR
+```
+## Stage 2: Top-k Models Averaging
+After training the model, we need to get the final model for testing and inference. In every epoch, the model checkpoint is saved, so we can choose the best model from them based on the validation loss or we can sort them and average the parameters of the top-k models to get the final model. We can use stage 2 to do this, and the code is shown below. Note: We only train one epoch for wavlmASR, thus the `avg_num` is set to 1.
+```bash
+ if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+     # avg n best model
+     avg.sh best exp/${ckpt}/checkpoints ${avg_num}
+ fi
+```
+The `avg.sh` is in the `../../../utils/` which is define in the `path.sh`.
+If you want to get the final model, you can use the script below to execute stage 0, stage 1, and stage 2:
+```bash
+bash run.sh --stage 0 --stop_stage 2
+```
+or you can run these scripts in the command line (only use CPU).
+
+```bash
+. ./path.sh
+. ./cmd.sh
+bash ./local/data.sh
+CUDA_VISIBLE_DEVICES= ./local/train.sh conf/wavlmASR.yaml wavlmASR
+avg.sh best exp/wavlmASR/checkpoints 1
+```
+## Stage 3: Model Testing
+The test stage is to evaluate the model performance. The code of test stage is shown below:
+```bash
+ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+     # test ckpt avg_n
+     CUDA_VISIBLE_DEVICES=0 ./local/test.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} || exit -1
+ fi
+```
+If you want to train a model and test it, you can use the script below to execute stage 0, stage 1, stage 2, and stage 3 :
+```bash
+bash run.sh --stage 0 --stop_stage 3
+```
+or you can run these scripts in the command line (only use CPU).
+```bash
+. ./path.sh
+. ./cmd.sh
+bash ./local/data.sh
+CUDA_VISIBLE_DEVICES= ./local/train.sh conf/wavlmASR.yaml wavlmASR
+avg.sh best exp/wavlmASR/checkpoints 1
+CUDA_VISIBLE_DEVICES= ./local/test.sh conf/wavlmASR.yaml conf/tuning/decode.yaml exp/wavlmASR/checkpoints/avg_1
+```
+## Pretrained Model
+You can get the pretrained wavlmASR from [this](../../../docs/source/released_model.md).
+
+using the `tar` scripts to unpack the model and then you can use the script to test the model.
+
+For example:
+```bash
+wget https://paddlespeech.bj.bcebos.com/wavlm/wavlmASR-base-100h-librispeech_ckpt_1.4.0.model.tar.gz
+tar xzvf wavlmASR-base-100h-librispeech_ckpt_1.4.0.model.tar.gz
+source path.sh
+# If you have process the data and get the manifest file， you can skip the following 2 steps
+bash local/data.sh --stage -1 --stop_stage -1
+bash local/data.sh --stage 2 --stop_stage 2
+CUDA_VISIBLE_DEVICES= ./local/test.sh conf/wavlmASR.yaml conf/tuning/decode.yaml exp/wavlmASR/checkpoints/avg_1
+```
+The performance of the released models are shown in [here](./RESULTS.md).
+
+
+## Stage 4: Single Audio File Inference
+In some situations, you want to use the trained model to do the inference for the single audio file. You can use stage 5. The code is shown below
+```bash
+ if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
+     # test a single .wav file
+     CUDA_VISIBLE_DEVICES=0 ./local/test_wav.sh ${conf_path} ${decode_conf_path} exp/${ckpt}/checkpoints/${avg_ckpt} ${audio_file} || exit -1
+ fi
+```
+you can train the model by yourself using ```bash run.sh --stage 0 --stop_stage 3```, or you can download the pretrained model through the script below:
+```bash
+wget https://paddlespeech.bj.bcebos.com/wavlm/wavlm_baseplus_libriclean_100h.tar.gz
+tar xzvf wavlm_baseplus_libriclean_100h.tar.gz
+```
+You can download the audio demo:
+```bash
+wget -nc https://paddlespeech.bj.bcebos.com/datasets/single_wav/en/demo_002_en.wav -P data/
+```
+You need to prepare an audio file or use the audio demo above, please confirm the sample rate of the audio is 16K. You can get the result of the audio demo by running the script below.
+```bash
+CUDA_VISIBLE_DEVICES= ./local/test_wav.sh conf/wavlmASR.yaml conf/tuning/decode.yaml exp/wavlmASR/checkpoints/avg_1 data/demo_002_en.wav
+```

examples/librispeech/asr5/RESULTS.md

@@ -0,0 +1,9 @@
+# LibriSpeech
+
+## WavLMASR
+Fintuning on train-clean-100
+train: Epoch 16, 4*A800-80G, batchsize: 16, accum_grad: 8
+
+| Model | Params | Config | Augmentation| Test set | Decode method | WER |  
+| --- | --- | --- | --- | --- | --- | --- |
+| WavLMASR | 326.16M | conf/wavlmasr.yaml | spec_aug | test-clean | greedy search | 0.0561 |  

examples/librispeech/asr5/avg.sh

@@ -0,0 +1,33 @@
+#! /usr/bin/env bash
+
+if [ $# != 3 ]; then
+    echo "usage: ${0} [best|latest] ckpt_dir avg_num"
+    exit -1
+fi
+
+avg_mode=${1} # best,latest
+ckpt_dir=${2}
+average_num=${3}
+decode_checkpoint=${ckpt_dir}/avg_${average_num}.pdparams
+
+if [ $avg_mode == best ];then
+    # best
+    python avg_model.py \
+    --dst_model ${decode_checkpoint} \
+    --ckpt_dir ${ckpt_dir}  \
+    --num ${average_num} \
+    --val_best
+else
+    # latest
+    python avg_model.py \
+    --dst_model ${decode_checkpoint} \
+    --ckpt_dir ${ckpt_dir}  \
+    --num ${average_num}
+fi
+
+if [ $? -ne 0 ]; then
+    echo "Failed in avg ckpt!"
+    exit 1
+fi
+
+exit 0

examples/librispeech/asr5/cmd.sh

@@ -0,0 +1,89 @@
+# ====== About run.pl, queue.pl, slurm.pl, and ssh.pl ======
+# Usage: <cmd>.pl [options] JOB=1:<nj> <log> <command...>
+# e.g.
+#   run.pl --mem 4G JOB=1:10 echo.JOB.log echo JOB
+#
+# Options:
+#   --time <time>: Limit the maximum time to execute.
+#   --mem <mem>: Limit the maximum memory usage.
+#   -–max-jobs-run <njob>: Limit the number parallel jobs. This is ignored for non-array jobs.
+#   --num-threads <ngpu>: Specify the number of CPU core.
+#   --gpu <ngpu>: Specify the number of GPU devices.
+#   --config: Change the configuration file from default.
+#
+# "JOB=1:10" is used for "array jobs" and it can control the number of parallel jobs.
+# The left string of "=", i.e. "JOB", is replaced by <N>(Nth job) in the command and the log file name,
+# e.g. "echo JOB" is changed to "echo 3" for the 3rd job and "echo 8" for 8th job respectively.
+# Note that the number must start with a positive number, so you can't use "JOB=0:10" for example.
+#
+# run.pl, queue.pl, slurm.pl, and ssh.pl have unified interface, not depending on its backend.
+# These options are mapping to specific options for each backend and
+# it is configured by "conf/queue.conf" and "conf/slurm.conf" by default.
+# If jobs failed, your configuration might be wrong for your environment.
+#
+#
+# The official documentation for run.pl, queue.pl, slurm.pl, and ssh.pl:
+#   "Parallelization in Kaldi": http://kaldi-asr.org/doc/queue.html
+# =========================================================~
+
+
+# Select the backend used by run.sh from "local", "sge", "slurm", or "ssh"
+cmd_backend='local'
+
+# Local machine, without any Job scheduling system
+if [ "${cmd_backend}" = local ]; then
+
+    # The other usage
+    export train_cmd="run.pl"
+    # Used for "*_train.py": "--gpu" is appended optionally by run.sh
+    export cuda_cmd="run.pl"
+    # Used for "*_recog.py"
+    export decode_cmd="run.pl"
+
+# "qsub" (SGE, Torque, PBS, etc.)
+elif [ "${cmd_backend}" = sge ]; then
+    # The default setting is written in conf/queue.conf.
+    # You must change "-q g.q" for the "queue" for your environment.
+    # To know the "queue" names, type "qhost -q"
+    # Note that to use "--gpu *", you have to setup "complex_value" for the system scheduler.
+
+    export train_cmd="queue.pl"
+    export cuda_cmd="queue.pl"
+    export decode_cmd="queue.pl"
+
+# "sbatch" (Slurm)
+elif [ "${cmd_backend}" = slurm ]; then
+    # The default setting is written in conf/slurm.conf.
+    # You must change "-p cpu" and "-p gpu" for the "partion" for your environment.
+    # To know the "partion" names, type "sinfo".
+    # You can use "--gpu * " by default for slurm and it is interpreted as "--gres gpu:*"
+    # The devices are allocated exclusively using "${CUDA_VISIBLE_DEVICES}".
+
+    export train_cmd="slurm.pl"
+    export cuda_cmd="slurm.pl"
+    export decode_cmd="slurm.pl"
+
+elif [ "${cmd_backend}" = ssh ]; then
+    # You have to create ".queue/machines" to specify the host to execute jobs.
+    # e.g. .queue/machines
+    #   host1
+    #   host2
+    #   host3
+    # Assuming you can login them without any password, i.e. You have to set ssh keys.
+
+    export train_cmd="ssh.pl"
+    export cuda_cmd="ssh.pl"
+    export decode_cmd="ssh.pl"
+
+# This is an example of specifying several unique options in the JHU CLSP cluster setup.
+# Users can modify/add their own command options according to their cluster environments.
+elif [ "${cmd_backend}" = jhu ]; then
+
+    export train_cmd="queue.pl --mem 2G"
+    export cuda_cmd="queue-freegpu.pl --mem 2G --gpu 1 --config conf/gpu.conf"
+    export decode_cmd="queue.pl --mem 4G"
+
+else
+    echo "$0: Error: Unknown cmd_backend=${cmd_backend}" 1>&2
+    return 1
+fi