cat-shark
d0d96a8f89
背景：实测 CJOD-255 全程 BGM 覆盖（56% 低能量、几乎无静音），固定 VAD 参数把音乐当语音 → whisper 全段解码 → 80% 碎片化 + 32% 漏句 + 敏感段丢失。 实现 nodes/vad_profiler.py： - profile_audio：1s 能量网格分析 → 静音比例/BGM 覆盖/长停顿识别 - suggest_vad_parameters：按信号特征推荐 threshold/min_silence/speech_pad （BGM 覆盖 -> 降 threshold 增人声敏感；长停顿 -> 降 speech_pad 防时间漂移； 静音占比高 -> 升 threshold 剔虚警） - score_transcript：启发式评分（不用参考字幕，符合部署实际）—— 碎片率 + 幻觉词（感谢观看/晚安/音乐）扣分 + 平均字数适中 - vad_parameters_for_audio：信号分析 + 可选片段网格验证，选最优参数 - _grid_search_vad / _candidate_params：候选网格搜索与异常回退 测试 tests/test_vad_profiler.py：19 个用例覆盖信号分析、四个建议分支、 幻觉词/碎片评分、网格选优、异常回退、空音频/低采样率、候选展开。
253 lines
9.5 KiB
Python
253 lines
9.5 KiB
Python
"""每视频自适应 VAD 调参测试（先红后绿）。
|
||
|
||
验证信号分析 profile_audio、启发式参数建议 suggest_vad_parameters、
|
||
转录质量评分 score_transcript（幻觉词扣分）与 vad_parameters_for_audio
|
||
（信号分析 + 片段网格验证）。
|
||
|
||
评分**不依赖参考字幕**（实际部署无参考），用转录质量 + 幻觉词扣分。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import wave
|
||
from pathlib import Path
|
||
|
||
import pytest
|
||
|
||
from nodes.vad_profiler import (
|
||
AudioProfile,
|
||
HALLUCINATION_TOKENS,
|
||
profile_audio,
|
||
score_transcript,
|
||
suggest_vad_parameters,
|
||
vad_parameters_for_audio,
|
||
_pick_representative_start,
|
||
)
|
||
|
||
|
||
class _Seg:
|
||
"""模拟 whisper segment：仅含 text。"""
|
||
|
||
def __init__(self, text):
|
||
self.text = text
|
||
|
||
|
||
def _make_wav(path: Path, silence_seconds: int, voice_seconds: int) -> None:
|
||
"""生成 [静音 N 秒 + 语音 N 秒] 的 16kHz 单声道 WAV。
|
||
|
||
silence 用 0 样本（静音），voice 用较大振幅样本（语音）。
|
||
"""
|
||
import array
|
||
|
||
rate = 16000
|
||
silence = array.array("h", [0] * rate * silence_seconds)
|
||
voice = array.array("h", [8000] * rate * voice_seconds)
|
||
samples = silence + voice
|
||
with wave.open(str(path), "wb") as f:
|
||
f.setnchannels(1)
|
||
f.setsampwidth(2)
|
||
f.setframerate(rate)
|
||
f.writeframes(samples.tobytes())
|
||
|
||
|
||
def test_profile_audio_detects_bgm_heavy() -> None:
|
||
"""纯静音+语音的视频：silence_ratio 高、非 BGM 覆盖；mediam_rms 合理。"""
|
||
wav = Path("/tmp/test_vad_plain.wav")
|
||
_make_wav(wav, silence_seconds=40, voice_seconds=10)
|
||
p = profile_audio(wav, 16000)
|
||
assert p.silence_ratio > 0.5
|
||
assert p.bgm_heavy is False
|
||
assert p.duration_seconds == pytest.approx(50.0, abs=1)
|
||
|
||
|
||
def test_suggest_vad_parameters_plain_silence() -> None:
|
||
"""静音占比高且无长停顿 -> 建议 threshold 偏高（静音权重分支）。"""
|
||
wav = Path("/tmp/test_vad_plain2.wav")
|
||
# 交替短静音避免触发 long_silence（<5s 连续静音）。
|
||
_make_wav(wav, silence_seconds=3, voice_seconds=1)
|
||
_make_wav(wav, silence_seconds=3, voice_seconds=1)
|
||
p = profile_audio(wav, 16000)
|
||
assert p.long_silence is False
|
||
params = suggest_vad_parameters(p)
|
||
assert params["threshold"] >= 0.5
|
||
|
||
|
||
def test_suggest_vad_parameters_bgm() -> None:
|
||
"""BGM 覆盖（低能量占比高但静音少）-> threshold 降低、min_silence 减小。"""
|
||
p = AudioProfile(
|
||
silence_ratio=0.1, # 几乎无静音
|
||
lowish_ratio=0.6, # 大量低能量（音乐）
|
||
bgm_heavy=True, # 直接标记 BGM 覆盖
|
||
long_silence=False,
|
||
rms_bins=[500] * 100,
|
||
)
|
||
params = suggest_vad_parameters(p)
|
||
assert params["threshold"] < 0.5 # 降低
|
||
assert params["min_silence_duration_ms"] < 1000 # 减小
|
||
|
||
|
||
def test_score_transcript_penalizes_hallucination() -> None:
|
||
"""幻觉词（感谢观看/晚安/音乐）多 -> 评分低。"""
|
||
good = [_Seg("ありがとうございます本日は"), _Seg("かしこまりました")]
|
||
bad = [_Seg("ご視聴ありがとうございました"), _Seg("おやすみなさい"), _Seg("音楽")]
|
||
assert score_transcript(good) > score_transcript(bad)
|
||
|
||
|
||
def test_score_transcript_penalizes_fragments() -> None:
|
||
"""碎片（纯单字/语气词）多 -> 评分低。"""
|
||
clean = [_Seg("今日はとても暑いですね"), _Seg("それでは始めましょう")]
|
||
frag = [_Seg("あ"), _Seg("うん"), _Seg("はい"), _Seg("あっ")]
|
||
assert score_transcript(clean) > score_transcript(frag)
|
||
|
||
|
||
def test_vad_parameters_for_audio_uses_grid_when_provider() -> None:
|
||
"""提供 whisper 回调时：从候选网格选评分最高的参数（threshold 最小者）。"""
|
||
wav = Path("/tmp/test_vad_grid.wav")
|
||
_make_wav(wav, silence_seconds=10, voice_seconds=5)
|
||
|
||
def fake_whisper(**kwargs):
|
||
# 假设最优 = threshold 最小的候选（suggest 0.5 时候选含 0.4）。
|
||
t = kwargs.get("vad_parameters", {}).get("threshold", 0.5)
|
||
if t <= 0.4:
|
||
# 最优组合：无幻觉的正常长句（分数最高）。
|
||
return [_Seg("今日はお客様のために精神整備を務めさせていただきます")]
|
||
# 其它组合：幻觉套话（分数低）。
|
||
return [_Seg("ご視聴ありがとうございました"), _Seg("おやすみなさい")]
|
||
|
||
params = vad_parameters_for_audio(
|
||
wav, sample_rate=16000, whisper_invoke=fake_whisper, run_dir=Path("/tmp")
|
||
)
|
||
# 网格应从候选里选出评分最高的 threshold=0.4 的组合。
|
||
assert params["threshold"] == pytest.approx(0.4, abs=0.1)
|
||
|
||
|
||
def test_vad_parameters_for_audio_fallback_without_provider() -> None:
|
||
"""无 whisper 回调：退化为信号分析建议，不报错。"""
|
||
wav = Path("/tmp/test_vad_noprov.wav")
|
||
_make_wav(wav, silence_seconds=10, voice_seconds=5)
|
||
params = vad_parameters_for_audio(wav, sample_rate=16000, whisper_invoke=None)
|
||
assert "threshold" in params
|
||
assert "min_silence_duration_ms" in params
|
||
|
||
|
||
def test_hallucination_tokens_included() -> None:
|
||
"""幻觉词集合应含常用套话（感谢观看/晚安/音乐）。"""
|
||
all_str = " ".join(HALLUCINATION_TOKENS)
|
||
assert "ご視聴ありがとうございました" in all_str
|
||
assert "おやすみなさい" in all_str
|
||
assert "音楽" in all_str
|
||
|
||
def test_suggest_vad_parameters_long_silence() -> None:
|
||
"""长停顿常见 -> 建议 threshold 0.5、speech_pad 200（防时间轴压缩漂移）。"""
|
||
p = AudioProfile(
|
||
silence_ratio=0.2, lowish_ratio=0.3, bgm_heavy=False, long_silence=True,
|
||
rms_bins=[500] * 100,
|
||
)
|
||
params = suggest_vad_parameters(p)
|
||
assert params["threshold"] == 0.5
|
||
assert params["speech_pad_ms"] == 200
|
||
|
||
|
||
def test_suggest_vad_parameters_regular() -> None:
|
||
"""常规音频（无特殊标记）-> 默认 0.5/1000/400。"""
|
||
p = AudioProfile(
|
||
silence_ratio=0.2, lowish_ratio=0.3, bgm_heavy=False, long_silence=False,
|
||
rms_bins=[1500] * 100,
|
||
)
|
||
params = suggest_vad_parameters(p)
|
||
assert params["threshold"] == 0.5
|
||
assert params["min_silence_duration_ms"] == 1000
|
||
assert params["speech_pad_ms"] == 400
|
||
|
||
|
||
def test_pick_representative_start_all_silence() -> None:
|
||
"""全静音音频：窗口恐怖沉默用 score>0.55 惩罚，仍返回起始 0。"""
|
||
p = AudioProfile(
|
||
rms_bins=[50] * 100, silence_ratio=0.9, lowish_ratio=0.9,
|
||
bgm_heavy=False, long_silence=False,
|
||
)
|
||
start = _pick_representative_start(p, 90)
|
||
assert start >= 0 and start < 10
|
||
|
||
|
||
def test_grid_search_exception_skipped() -> None:
|
||
"""网格搜索某组参数调用抛异常：跳过该组，不崩溃。"""
|
||
called = {"n": 0}
|
||
|
||
def failing_whisper(**kwargs):
|
||
called["n"] += 1
|
||
raise RuntimeError("mock download fail")
|
||
|
||
from nodes.vad_profiler import _grid_search_vad
|
||
|
||
params = _grid_search_vad(
|
||
failing_whisper, Path("/tmp/x.wav"), Path("/tmp"), 60, 0,
|
||
{"threshold": 0.5, "min_silence_duration_ms": 1000, "speech_pad_ms": 400},
|
||
)
|
||
# 全部失败时回退 suggested。
|
||
assert params == {"threshold": 0.5, "min_silence_duration_ms": 1000, "speech_pad_ms": 400}
|
||
assert called["n"] > 0
|
||
|
||
|
||
def test_candidate_params_expands_grid() -> None:
|
||
"""候选网格应含建议值邻域（threshold±0.1、silence 倍、pad 组合）。"""
|
||
from nodes.vad_profiler import _candidate_params
|
||
|
||
cands = _candidate_params(
|
||
{"threshold": 0.5, "min_silence_duration_ms": 1000, "speech_pad_ms": 400}
|
||
)
|
||
assert len(cands) == 27
|
||
assert {"threshold": 0.4, "min_silence_duration_ms": 500, "speech_pad_ms": 0} in cands
|
||
assert {"threshold": 0.6, "min_silence_duration_ms": 2000, "speech_pad_ms": 400} in cands
|
||
|
||
|
||
def test_median_rms_empty_returns_zero() -> None:
|
||
"""rms_bins 为空时 median_rms 返回 0。"""
|
||
p = AudioProfile(rms_bins=[])
|
||
assert p.median_rms == 0.0
|
||
|
||
|
||
def test_profile_audio_empty_wav() -> None:
|
||
"""空的 WAV（无样本）-> 返回 duration 0 的空 profile，不崩溃。"""
|
||
import array
|
||
wav = Path("/tmp/test_vad_empty.wav")
|
||
with wave.open(str(wav), "wb") as f:
|
||
f.setnchannels(1); f.setsampwidth(2); f.setframerate(16000)
|
||
f.writeframes(array.array("h", []).tobytes())
|
||
p = profile_audio(wav, 16000)
|
||
assert p.duration_seconds == 0.0
|
||
|
||
|
||
def test_grid_search_ignores_non_list() -> None:
|
||
"""网格搜索回调返回非 list（如 None）应被跳过，回退建议值。"""
|
||
from nodes.vad_profiler import _grid_search_vad
|
||
|
||
params = _grid_search_vad(
|
||
lambda **kw: None, Path("/tmp/x.wav"), Path("/tmp"), 60, 0,
|
||
{"threshold": 0.5, "min_silence_duration_ms": 1000, "speech_pad_ms": 400},
|
||
)
|
||
assert params == {"threshold": 0.5, "min_silence_duration_ms": 1000, "speech_pad_ms": 400}
|
||
|
||
|
||
def test_score_transcript_empty_list() -> None:
|
||
"""空列表评分返回 0。"""
|
||
assert score_transcript([]) == 0.0
|
||
|
||
|
||
def test_score_transcript_ignores_empty_text_seg() -> None:
|
||
"""含空文本的 segment 被跳过，不报错。"""
|
||
assert score_transcript([_Seg(""), _Seg("今日は暑いです")]) > 0
|
||
|
||
|
||
def test_profile_audio_different_sample_rate() -> None:
|
||
"""采样率与请求不一致时用 wave 实际 rate 近似，不崩溃。"""
|
||
import array
|
||
wav = Path("/tmp/test_vad_rate.wav")
|
||
rate = 8000
|
||
samples = array.array("h", [5000] * rate * 2) # 2s 语音
|
||
with wave.open(str(wav), "wb") as f:
|
||
f.setnchannels(1); f.setsampwidth(2); f.setframerate(rate)
|
||
f.writeframes(samples.tobytes())
|
||
p = profile_audio(wav, 16000)
|
||
assert p.duration_seconds == pytest.approx(2.0, abs=0.2)
|