| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Regression test for chunk timestamp offset drift in _preprocess_speech_to_text.""" |
| 4 | |
| 5 | from __future__ import annotations |
| 6 | |
| 7 | from unittest.mock import AsyncMock, MagicMock, patch |
| 8 | |
| 9 | import numpy as np |
| 10 | import pytest |
| 11 | |
| 12 | from vllm.config.speech_to_text import SpeechToTextConfig |
| 13 | from vllm.entrypoints.speech_to_text.base.serving import SpeechToTextBaseServing |
| 14 | |
| 15 | SR = 16_000 |
| 16 | _PATCH = "vllm.entrypoints.speech_to_text.base.serving" |
| 17 | |
| 18 | |
| 19 | @pytest.mark.asyncio |
| 20 | async def test_chunk_offsets_are_cumulative_not_nominal(): |
| 21 | """chunk_start_offsets must be cumulative actual chunk lengths, not the old approach |
| 22 | of 'idx * max_audio_clip_s'. When split_audio places a boundary before the |
| 23 | nominal 30 s mark, the old formula drifts; the fixed formula stays exact. |
| 24 | """ |
| 25 | # Chunks shorter than exactly 30 s, as split_audio produces when a quiet |
| 26 | # region falls inside the 1 s overlap window before the nominal boundary. |
| 27 | chunk_lengths = [int(29.5 * SR), int(29.7 * SR), int(5.0 * SR)] |
| 28 | chunks = [np.zeros(n, dtype=np.float32) for n in chunk_lengths] |
| 29 | |
| 30 | duration = sum(chunk_lengths) / SR |
| 31 | |
| 32 | expected_offsets = [0.0, 29.5, 29.5 + 29.7] # cumulative seconds |
| 33 | wrong_offsets = [0.0, 30.0, 60.0] # what the old bug produced |
| 34 | |
| 35 | serving = SpeechToTextBaseServing.__new__(SpeechToTextBaseServing) |
| 36 | serving._decode_and_chunk_speech_async = AsyncMock(return_value=(chunks, duration)) |
| 37 | serving.asr_config = SpeechToTextConfig( |
| 38 | sample_rate=float(SR), |
| 39 | max_audio_clip_s=30, |
| 40 | overlap_chunk_second=1, |
| 41 | min_energy_split_window_size=1600, |
| 42 | ) |
| 43 | serving.model_cls = MagicMock() |
| 44 | serving.model_cls.validate_language.side_effect = lambda lang: lang |
| 45 | serving.model_cls.supports_explicit_language_detection = False |
| 46 | serving.model_cls.get_generation_prompt.return_value = {} |
| 47 | serving.model_config = MagicMock() |
| 48 | serving.task_type = "transcribe" |
| 49 | serving.renderer = MagicMock() |
| 50 | serving.renderer.render_cmpl_async = AsyncMock( |
| 51 | return_value=[MagicMock()] * len(chunks) |
| 52 | ) |
| 53 | |
| 54 | request = MagicMock() |
| 55 | request.language = "en" |
| 56 | request.to_language = None |
| 57 | request.response_format = "json" |
| 58 | request.build_stt_params.return_value = MagicMock() |
| 59 | |
| 60 | with patch(f"{_PATCH}.parse_model_prompt", return_value=MagicMock()): |
| 61 | _, _, offsets = await serving._preprocess_speech_to_text( |
| 62 | request=request, |
| 63 | audio_data=b"\x00", |
| 64 | request_id="test", |
| 65 | ) |
| 66 | |
| 67 | assert offsets == pytest.approx(expected_offsets, abs=1e-6) |
| 68 | assert offsets != pytest.approx(wrong_offsets, abs=1e-6) |