[ANDROIDHASSAN]

Problem

With --batch_inference, a chunk where VAD finds no speech is still decoded when it lands in a batch of 2 or more requests.

prepare_features only drops the non-speech audio if speech_chunks:. When get_speech_timestamps returns [], the full unfiltered chunk is kept, so speech_duration is never 0.0. The "no speech detected" early return in _process_multi is therefore never reached for VAD requests, and the whole chunk goes through encode + generate.

The non-batched path (transcriber.transcribe, also used by the batch worker for a single request) collects chunks unconditionally. collect_chunks(audio, []) returns one empty array, so transcribe returns None, None and nothing is emitted. As a result, the same audio gives different output depending on how many sessions happen to share the batch window.

Fix

prepare_features now does what transcriber.transcribe does:

speech_chunks = get_speech_timestamps(audio, vad_opts)
audio_chunks, _ = collect_chunks(audio, speech_chunks)
audio = np.concatenate(audio_chunks, axis=0)

With no speech, the audio is empty and speech_duration is 0.0. The existing early return then sends back [] without touching the GPU.

Tests

• New test test_vad_without_speech_skips_decoding. It runs two use_vad=True requests where VAD returns [] and checks that encode is not called and both results are [].
  • It fails on main with Expected 'encode' to not have been called. Called 1 times.
  • It passes with this change.
  • Restoring the old guard makes it fail again.
• pytest tests/test_batch_inference.py tests/test_batch_timeout.py: 24 passed.
• Across all of tests/, the results are the same as on main apart from the new test. In my venv, main gives 109 passed and this branch gives 110. The same 2 failures and the same 8 collection errors show up on both, from optional deps I did not install (openvino, etc.).
• flake8 with the CI's --select=E9,F63,F7,F82 reports 0 on both files.

Real model check, run on CPU with tiny int8. assets/jfk.flac (first 10 s) is submitted at falling gains as 5 batched use_vad=True requests, and each result is compared with transcriber.transcribe(..., vad_filter=True) on the same audio:

gain	VAD chunks	batch, main	batch, this PR	transcribe
1.0 – 0.01	1–2	text	text (unchanged)	text
0.003	0	['And so my fellow Americans ask not what your country can do for you ask what you can do for your']	[]	[]

All rows where VAD finds speech give the same output before and after the change.

Scope

One function in whisper_live/batch_inference.py and one test. The single-request path and the non-batched path are untouched.