Whisper が無音区間に『ご視聴ありがとうございました』を書き足していたとき — 日本語文字起こしの幻聴を計測して潰す運用メモ

def filter_hallucinations(result):
    kept, removed = [], []
    for seg in result.segments:
        is_silence = seg.no_speech_prob > 0.6
        is_low_conf = seg.avg_logprob < -1.0
        # 無音確率が高く、かつ確信度も低いものだけ落とす
        if is_silence and is_low_conf:
            removed.append(seg.text.strip())
            continue
        kept.append(seg.text.strip())
 
    cleaned = " ".join(t for t in kept if t)
    return cleaned, removed

import webrtcvad
import wave
import contextlib
 
def has_speech(frame_bytes, vad, sample_rate=16000):
    return vad.is_speech(frame_bytes, sample_rate)
 
def strip_silence(pcm_path, aggressiveness=2):
    # aggressiveness は 0（緩い）〜3（厳しい）。会議音声は 2 が扱いやすい
    vad = webrtcvad.Vad(aggressiveness)
    speech_frames = []
    with contextlib.closing(wave.open(pcm_path, "rb")) as wf:
        sample_rate = wf.getframerate()
        frame_ms = 30
        frame_bytes = int(sample_rate * frame_ms / 1000) * 2  # 16bit=2byte
        while True:
            frame = wf.readframes(frame_bytes // 2)
            if len(frame) < frame_bytes:
                break
            if has_speech(frame, vad, sample_rate):
                speech_frames.append(frame)
 
    return b"".join(speech_frames)

import anthropic, json
 
client = anthropic.Anthropic(api_key="YOUR_API_KEY")
 
def validate_transcript(segments):
    # segments: [{"start": 12.4, "text": "..."}] のリスト
    numbered = "\n".join(f"[{i}] {s['text']}" for i, s in enumerate(segments))
 
    resp = client.messages.create(
        model="claude-sonnet-5",
        max_tokens=1024,
        tools=[{
            "name": "report_hallucinations",
            "description": "文脈から浮いた挿入セグメントの番号を返す",
            "input_schema": {
                "type": "object",
                "properties": {
                    "suspect_indices": {
                        "type": "array",
                        "items": {"type": "integer"},
                        "description": "前後と繋がらない、または定型句のみのセグメント番号",
                    }
                },
                "required": ["suspect_indices"],
            },
        }],
        tool_choice={"type": "tool", "name": "report_hallucinations"},
        messages=[{
            "role": "user",
            "content": (
                "以下は会議音声の文字起こしです。前後の文脈と繋がらず、"
                "「ご視聴ありがとうございました」等の配信定型句だけのセグメントを"
                "幻聴として検出してください。本物の発話は残してください。\n\n" + numbered
            ),
        }],
    )
 
    for block in resp.content:
        if block.type == "tool_use":
            return block.input["suspect_indices"]
    return []

BLOCKLIST = ["ご視聴ありがとうございました", "チャンネル登録"]
 
def blocklist_filter(segments):
    result = []
    for s in segments:
        text = s["text"].strip()
        # 定型句「のみ」で構成され、かつ無音確率が高い場合だけ落とす
        only_boilerplate = any(text == b for b in BLOCKLIST)
        if only_boilerplate and s.get("no_speech_prob", 0) > 0.6:
            continue
        result.append(s)
    return result