update timestamp doc

egs_modelscope/tp/TEMPLATE/README.md

@@ -0,0 +1,102 @@
+# TIMESTAMP PREDICTION
+
+## Inference
+
+### Quick start
+#### [Use TP-Aligner Model Simply](https://modelscope.cn/models/damo/speech_timestamp_prediction-v1-16k-offline/summary)
+```python
+from modelscope.pipelines import pipeline
+from modelscope.utils.constant import Tasks
+
+inference_pipline = pipeline(
+    task=Tasks.speech_timestamp,
+    model='damo/speech_timestamp_prediction-v1-16k-offline',
+    output_dir=None)
+
+rec_result = inference_pipline(
+    audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_timestamps.wav',
+    text_in='一 个 东 太 平 洋 国 家 为 什 么 跑 到 西 太 平 洋 来 了 呢',)
+print(rec_result)
+```
+
+Timestamp pipeline can also be used after ASR pipeline to compose complete ASR function, ref to [demo](https://github.com/alibaba-damo-academy/FunASR/discussions/246).
+
+
+
+#### API-reference
+##### Define pipeline
+- `task`: `Tasks.speech_timestamp`
+- `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
+- `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 
+- `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 speech to predict, which could be: 
+  - wav_path, `e.g.`: asr_example.wav (wav in local or url), 
+  - wav.scp, kaldi style wav list (`wav_id wav_path`), `e.g.`: 
+    ```text
+    asr_example1  ./audios/asr_example1.wav
+    asr_example2  ./audios/asr_example2.wav
+    ```
+  In this case of `wav.scp` input, `output_dir` must be set to save the output results
+- `text_in`: the input text to predict， splited by blank, which could be:
+  - text string, `e.g.`: `今 天 天 气 怎 么 样`
+  - text.scp, kaldi style text file (`wav_id transcription`), `e.g.`:
+    ```text
+    asr_example1 今 天 天 气 怎 么 样
+    asr_example2 欢 迎 体 验 达 摩 院 语 音 识 别 模 型
+    ```
+- `audio_fs`: audio sampling rate, only set when audio_in is pcm audio
+- `output_dir`: None (Default), the output path of results if set, containing
+  - output_dir/timestamp_prediction/tp_sync, timestamp in second containing silence periods, `wav_id# token1 start_time end_time;`, `e.g.`:
+    ```text
+    test_wav1# <sil> 0.000 0.500;温 0.500 0.680;州 0.680 0.840;化 0.840 1.040;工 1.040 1.280;仓 1.280 1.520;<sil> 1.520 1.680;库 1.680 1.920;<sil> 1.920 2.160;起 2.160 2.380;火 2.380 2.580;殃 2.580 2.760;及 2.760 2.920;附 2.920 3.100;近 3.100 3.340;<sil> 3.340 3.400;河 3.400 3.640;<sil> 3.640 3.700;流 3.700 3.940;<sil> 3.940 4.240;大 4.240 4.400;量 4.400 4.520;死 4.520 4.680;鱼 4.680 4.920;<sil> 4.920 4.940;漂 4.940 5.120;浮 5.120 5.300;河 5.300 5.500;面 5.500 5.900;<sil> 5.900 6.240;
+    ```
+  - output_dir/timestamp_prediction/tp_time, timestamp list in ms of same length as input text without silence `wav_id# [[start_time, end_time],]`, `e.g.`:
+    ```text
+    test_wav1# [[500, 680], [680, 840], [840, 1040], [1040, 1280], [1280, 1520], [1680, 1920], [2160, 2380], [2380, 2580], [2580, 2760], [2760, 2920], [2920, 3100], [3100, 3340], [3400, 3640], [3700, 3940], [4240, 4400], [4400, 4520], [4520, 4680], [4680, 4920], [4940, 5120], [5120, 5300], [5300, 5500], [5500, 5900]]
+    ```
+
+### Inference with multi-thread CPUs or multi GPUs
+FunASR also offer recipes [egs_modelscope/vad/TEMPLATE/infer.sh](https://github.com/alibaba-damo-academy/FunASR/blob/main/egs_modelscope/vad/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 **must** include `wav.scp` and `text.scp`
+    - `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
+
+- Decode with multi GPUs:
+```shell
+    bash infer.sh \
+    --model "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch" \
+    --data_dir "./data/test" \
+    --output_dir "./results" \
+    --batch_size 64 \
+    --gpu_inference true \
+    --gpuid_list "0,1"
+```
+- Decode with multi-thread CPUs:
+```shell
+    bash infer.sh \
+    --model "damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch" \
+    --data_dir "./data/test" \
+    --output_dir "./results" \
+    --gpu_inference false \
+    --njob 64
+```
+
+## Finetune with pipeline
+
+### Quick start
+
+### Finetune with your data
+
+## Inference with your finetuned model
+

egs_modelscope/tp/TEMPLATE/infer.py

@@ -0,0 +1 @@
+../speech_timestamp_prediction-v1-16k-offline/infer.py

egs_modelscope/tp/TEMPLATE/infer.sh

@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+
+set -e
+set -u
+set -o pipefail
+
+stage=1
+stop_stage=2
+model="damo/speech_timestamp_prediction-v1-16k-offline"
+data_dir="./data/test"
+output_dir="./results"
+batch_size=1
+gpu_inference=true    # whether to perform gpu decoding
+gpuid_list="0,1"    # set gpus, e.g., gpuid_list="0,1"
+njob=4    # the number of jobs for CPU decoding, if gpu_inference=false, use CPU decoding, please set njob
+checkpoint_dir=
+checkpoint_name="valid.cer_ctc.ave.pb"
+
+. utils/parse_options.sh || exit 1;
+
+if ${gpu_inference} == "true"; then
+    nj=$(echo $gpuid_list | awk -F "," '{print NF}')
+else
+    nj=$njob
+    batch_size=1
+    gpuid_list=""
+    for JOB in $(seq ${nj}); do
+        gpuid_list=$gpuid_list"-1,"
+    done
+fi
+
+mkdir -p $output_dir/split
+split_scps=""
+split_texts=""
+for JOB in $(seq ${nj}); do
+    split_scps="$split_scps $output_dir/split/wav.$JOB.scp"
+    split_texts="$split_texts $output_dir/split/text.$JOB.scp"
+done
+perl utils/split_scp.pl ${data_dir}/wav.scp ${split_scps}
+perl utils/split_scp.pl ${data_dir}/text.scp ${split_texts}
+
+if [ -n "${checkpoint_dir}" ]; then
+  python utils/prepare_checkpoint.py ${model} ${checkpoint_dir} ${checkpoint_name}
+  model=${checkpoint_dir}/${model}
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ];then
+    echo "Decoding ..."
+    gpuid_list_array=(${gpuid_list//,/ })
+    for JOB in $(seq ${nj}); do
+        {
+        id=$((JOB-1))
+        gpuid=${gpuid_list_array[$id]}
+        mkdir -p ${output_dir}/output.$JOB
+        python infer.py \
+            --model ${model} \
+            --audio_in ${output_dir}/split/wav.$JOB.scp \
+            --text_in ${output_dir}/split/text.$JOB.scp \
+            --output_dir ${output_dir}/output.$JOB \
+            --batch_size ${batch_size} \
+            --gpuid ${gpuid}
+        }&
+    done
+    wait
+
+    mkdir -p ${output_dir}/timestamp_prediction
+    for f in tp_sync tp_time; do
+        if [ -f "${output_dir}/output.1/timestamp_prediction/${f}" ]; then
+          for i in $(seq "${nj}"); do
+              cat "${output_dir}/output.${i}/timestamp_prediction/${f}"
+          done | sort -k1 >"${output_dir}/timestamp_prediction/${f}"
+        fi
+    done
+fi
+

egs_modelscope/tp/TEMPLATE/utils

@@ -0,0 +1 @@
+../../vad/TEMPLATE/utils

egs_modelscope/tp/speech_timestamp_prediction-v1-16k-offline/infer.py

@@ -1,12 +1,28 @@
+import os
+import argparse
 from modelscope.pipelines import pipeline
 from modelscope.utils.constant import Tasks
 
-inference_pipline = pipeline(
-    task=Tasks.speech_timestamp,
-    model='damo/speech_timestamp_prediction-v1-16k-offline',
-    output_dir='./tmp')
+def modelscope_infer(args):
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpuid)
+    inference_pipeline = pipeline(
+        task=Tasks.speech_timestamp,
+        model=args.model,
+        output_dir=args.output_dir,
+        batch_size=args.batch_size,
+    )
+    if args.output_dir is not None:
+        inference_pipeline(audio_in=args.audio_in, text_in=args.text_in)
+    else:
+        print(inference_pipeline(audio_in=args.audio_in, text_in=args.text_in))
 
-rec_result = inference_pipline(
-    audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_timestamps.wav',
-    text_in='一 个 东 太 平 洋 国 家 为 什 么 跑 到 西 太 平 洋 来 了 呢')
-print(rec_result)
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model', type=str, default="damo/speech_timestamp_prediction-v1-16k-offline")
+    parser.add_argument('--audio_in', type=str, default="https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_timestamps.wav")
+    parser.add_argument('--text_in', type=str, default="一 个 东 太 平 洋 国 家 为 什 么 跑 到 西 太 平 洋 来 了 呢")
+    parser.add_argument('--output_dir', type=str, default="./results/")
+    parser.add_argument('--batch_size', type=int, default=1)
+    parser.add_argument('--gpuid', type=str, default="0")
+    args = parser.parse_args()
+    modelscope_infer(args)

egs_modelscope/vad/TEMPLATE/README.md

@@ -1,7 +1,7 @@
 # Voice Activity Detection
 
 > **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 model of FSMN-VAD as example to demonstrate the usage.
+> 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 finetune. Here we take the model of FSMN-VAD as example to demonstrate the usage.
 
 ## Inference
 
@@ -57,7 +57,7 @@ Full code of demo, please ref to [demo](https://github.com/alibaba-damo-academy/
   - 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

funasr/bin/tp_inference.py

@@ -222,6 +222,13 @@ def inference_modelscope(
         split_with_space=split_with_space,
         seg_dict_file=seg_dict_file,
     )
+
+    if output_dir is not None:
+        writer = DatadirWriter(output_dir)
+        tp_writer = writer[f"timestamp_prediction"]
+        # ibest_writer["token_list"][""] = " ".join(speech2text.asr_train_args.token_list)
+    else:
+        tp_writer = None
     
     def _forward(
             data_path_and_name_and_type,
@@ -230,7 +237,14 @@ def inference_modelscope(
             fs: dict = None,
             param_dict: dict = None,
             **kwargs
-    ):
+    ):  
+        output_path = output_dir_v2 if output_dir_v2 is not None else output_dir
+        writer = None
+        if output_path is not None:
+            writer = DatadirWriter(output_path)
+            tp_writer = writer[f"timestamp_prediction"]
+        else:
+            tp_writer = None
         # 3. Build data-iterator
         if data_path_and_name_and_type is None and raw_inputs is not None:
             if isinstance(raw_inputs, torch.Tensor):
@@ -268,6 +282,9 @@ def inference_modelscope(
                 ts_str, ts_list = ts_prediction_lfr6_standard(us_alphas[batch_id], us_cif_peak[batch_id], token, force_time_shift=-3.0)
                 logging.warning(ts_str)
                 item = {'key': key, 'value': ts_str, 'timestamp':ts_list}
+                if tp_writer is not None:
+                    tp_writer["tp_sync"][key+'#'] = ts_str
+                    tp_writer["tp_time"][key+'#'] = str(ts_list)
                 tp_result_list.append(item)
         return tp_result_list