From 2e36e738ca39afc8d02f3d11013bd12f937cc874 Mon Sep 17 00:00:00 2001
From: zhaomingwork <61895407+zhaomingwork@users.noreply.github.com>
Date: 星期三, 08 十一月 2023 09:22:06 +0800
Subject: [PATCH] fix bug for h5 hotwords (#1067)

---
 egs_modelscope/asr/TEMPLATE/README.md |   28 ++++++++++++++++++++++++++++
 1 files changed, 28 insertions(+), 0 deletions(-)

diff --git a/egs_modelscope/asr/TEMPLATE/README.md b/egs_modelscope/asr/TEMPLATE/README.md
index ac73950..bf6d30b 100644
--- a/egs_modelscope/asr/TEMPLATE/README.md
+++ b/egs_modelscope/asr/TEMPLATE/README.md
@@ -21,6 +21,34 @@
 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-long Model](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary)
+```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-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch',
+    vad_model='damo/speech_fsmn_vad_zh-cn-16k-common-pytorch',
+    #punc_model='damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch',
+    punc_model='damo/punc_ct-transformer_cn-en-common-vocab471067-large',
+)
+
+rec_result = inference_pipeline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/vad_example.wav', 
+                                batch_size_token=5000, batch_size_token_threshold_s=40, max_single_segment_time=6000)
+print(rec_result)
+```
+
+Where, 
+- `batch_size_token` refs to dynamic batch_size and the total tokens of batch is `batch_size_token`, 1 token = 60 ms. 
+- `batch_size_token_threshold_s`: The batch_size is set to 1, when the audio duration exceeds the threshold value of `batch_size_token_threshold_s`, specified in `s`.
+- `max_single_segment_time`: The maximum length for audio segmentation in VAD, specified in `ms`.
+
+Suggestion: When encountering OOM (Out of Memory) issues with long audio inputs, as the GPU memory usage increases with the square of the audio duration, there are three possible scenarios:
+- a) In the initial inference stage, GPU memory usage primarily depends on `batch_size_token`. Reducing this value appropriately can help reduce memory usage. 
+- b) In the middle of the inference process, when encountering long audio segments segmented by VAD, if the total number of tokens is still smaller than `batch_size_token` but OOM issues persist, reducing `batch_size_token_threshold_s` can help. If the threshold is exceeded, forcing the batch size to 1 can be considered. 
+- c) Towards the end of the inference process, when encountering long audio segments segmented by VAD and the total number of tokens is smaller than `batch_size_token` but exceeds the threshold `batch_size_token_threshold_s`, forcing the batch size to 1 may still result in OOM errors. In such cases, reducing `max_single_segment_time` can be considered to shorten the duration of audio segments generated by VAD.
+
 #### [Paraformer-online Model](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online/summary)
 ##### Streaming Decoding
 ```python

--
Gitblit v1.9.1