| | |
|---|
| | | 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 |
|---|
