fix(auto): do not divide by zero speech time in AutoModel.inference - #3896

Summary

AutoModel.inference computes time_escape / batch_data_time for its rtf statistic after the model has already returned its result. fsmn-vad reports batch_data_time == 0 when a call produces no feature frame, so the statistic raised ZeroDivisionError and the result was lost. This happens for one-shot inputs shorter than 400 samples (25 ms at 16 kHz) and for streaming VAD with chunk_size=10 or 20 ms, where the first chunk produces no frame. The progress-bar rtf_avg description has the same division. Both now report nan when no speech time was reported. Models that report -1 (no timing) keep the previous value (code reading).

Type of change

• Bug fix

Validation

• python -m compileall funasr examples tests
• python -m pytest tests/test_auto_model_zero_speech_time.py -q: 2 failed on main (ZeroDivisionError at auto_model.py:986), 2 passed with the fix.
• fsmn-vad (iic/speech_fsmn_vad_zh-cn-16k-common-pytorch), CPU:
  • generate(input=wav[:n]): main raises for n <= 399 and works from n = 400 (bisected); with the fix n = 100 and 399 return value: [].
  • Streaming loop (cache, is_final) over the first 20 s of example/vad_example.wav: on main chunk_size=10 and 20 raise at the first chunk, 30 ms and larger work. With the fix 10 and 20 ms give the same events as 200 ms ([[70, -1], [-1, 19980]]).
  • Batch generate(input=[wav_20s, wav[:100]]): main raises and returns nothing; with the fix the 20 s input gets [[70, 19980]].
  • AutoModel(model=..., vad_model=<fsmn-vad>) on 100 samples: main raises in the VAD step; with the fix it returns {'text': '', 'timestamp': []}. VAD finds no segment, so the second model is never called (fsmn-vad stood in for the ASR model).
• Existing offline suites (test_load_audio_bytes, test_pcm_input_format, test_merge_vad_length, test_wav_frontend_no_lfr, test_fsmn_vad_*, test_dynamic_streaming_vad, test_punc_model_none, test_silero_vad_adapter, test_auto_model_ncpu): same results as on main.

User impact

Streaming VAD with 10 or 20 ms chunks (common audio-callback sizes), very short clips, and batches or VAD+ASR pipelines that contain such a clip no longer crash. A too-short input returns an empty result, as funasr_onnx does since #3822.

Notes for reviewers

• Same class as time_speech_total_all_samples除零异常 #1294 (fixed in inference_with_vad with a 1e-6 floor). nan is used here because a per-batch rtf with no speech time is undefined.
• A (samples, channels) array also reaches this division today: fsmn-vad sees no frame and raises the same ZeroDivisionError. With this change alone such input returns an empty result instead of raising. Open fix(audio): average the channels of 2-D waveform arrays and tensors #3891 averages 2-D arrays to mono; with it, such input returns the normal segments.
• Checked on CPU with fsmn-vad only; no ASR weights or GPU.