Defects Missed in Transcription — AI Speaks After 0.5-Second Silence

📝 Originally published (in Japanese) at

[forge.workstyle.tech].

I used to perform quality control (QC) for TTS models using this process:

I created 12 voices and passed all of them through this QC. Whisper got 4/4 accuracy, no trailing elongation, and sound pressure was within normal range. I reported 100% pass rate.

Later, when I rechecked from a different angle, 4 of them still had defects. These were invisible to STT-based inspection due to fundamental limitations.

The first clue came when I received this report:

ご覧ください。   Total: 1.61s  Body: 0.88s → Silence: 0.48s → 【0.16s utterance】
こちらです。     Total: 1.65s  Body: 0.72s → Silence: 0.56s → 【0.28s utterance】

After finishing the script, there’s a full 0.5-second silence followed by a 0.1–0.3 second utterance. This isn’t trailing resonance—the model is producing sounds not in the script (the root cause was training corpus contamination: "3 characters" allowed by the quality gate became verbal tics).

The reason my initial inspection missed this is simple: Whisper dropped these sounds.

ご覧ください。   → STT: "ご覧くださいああ"     ← barely caught
こちらです。     → STT: "こちらです"           ← completely dropped

A 0.28-second utterance doesn’t appear in the transcription at all. Short sounds that aren’t meaningful words may not appear in STT output. As long as you’re only looking at transcriptions, this defect doesn’t exist.

I even concluded, “STT got 0/6, so no extra sounds,” mistaking the blind spot of my measurement method for a property of the target.

What’s actually being output appears in the waveform. By extracting voiced blocks from the RMS envelope and examining their sequence, we can detect these artifacts.

def segments(wav_bytes, thr_ratio=0.06):
    """Returns [(start_sec, end_sec), ...] of voiced blocks"""
    w = wave.open(io.BytesIO(wav_bytes)); sr = w.getframerate()
    x = np.frombuffer(w.readframes(w.getnframes()), dtype=np.int16) / 32768

    W, H = int(sr * 0.020), int(sr * 0.010)          # 20ms window / 10ms hop
    rms = np.array([np.sqrt(np.mean(x[i*H:i*H+W]**2))
                    for i in range(max(0, (len(x)-W)//H))])

    act = rms > max(rms.max() * thr_ratio, 0.004)

    segs, s = [], None
    for i, a in enumerate(act):
        if a and s is None:
            s = i
        elif not a and s is not None:
            if (i - s) * 0.010 >= 0.03:              # Ignore blocks <30ms
                segs.append((s * 0.010, i * 0.010))
            s = None
    if s is not None:
        segs.append((s * 0.010, len(act) * 0.010))
    return segs

The max(rms.max() * 0.06, 0.004)

part is subtly critical.

Relative threshold alone fails for low-volume voices. If the overall volume is quiet, the maximum value is small, causing noise floor to be misclassified as voiced.

Absolute threshold alone fails for high-volume voices. Breathing or lip smacks get classified as voiced.

The 12 voices had sound pressure ranging from −13.3 to −18.8 dB, so neither threshold alone could work across all voices.

Discarding blocks under 30ms is also necessary. Without this, lip noise or quantization noise appears as many tiny blocks, breaking downstream logic.

Once voiced blocks are extracted, we check: “Is there sufficient silence before the final block, and does that block have sufficient duration?”

GAP_MIN  = 0.25      # Silence this long or more indicates a separate utterance
TAIL_MIN = 0.06      # Duration this long or more indicates an artifact

def has_trailing_artifact(wav):
    segs = segments(wav)
    if len(segs) < 2:
        return None                      # No artifact if only one block
    gap  = segs[-1][0] - segs[-2][1]     # Silence before last block
    tail = segs[-1][1] - segs[-1][0]     # Duration of last block
    if gap >= GAP_MIN and tail >= TAIL_MIN:
        return (gap, tail)
    return None

GAP_MIN=0.25

separates natural trailing resonance or s from clearly separated utterances. Measured artifacts had silence gaps of 0.26–0.91 seconds, so 0.25 is sufficient.

TAIL_MIN=0.06

avoids catching fade-out tails. Measured artifacts were 0.07–0.36 seconds long.

In my first scan, all 12 models triggered the detector. The probe sentence contained this:

では、始めます。
  Block[0] = "では"
  Block[1] = "始めます"   ← 0.40s silence followed by 0.75s duration

This was a after a comma. After “では、” there’s a gap, then “始めます。” follows. The final block is part of the script itself, yet it perfectly matches the detection condition (gap + subsequent utterance).

The condition “there’s utterance after the final gap” will always produce false positives for sentences containing commas. That’s because it doesn’t consider script structure.

There are two fixes:

Limit probes to single sentences. If you exclude sentences with commas, any utterance after the body can be definitively identified as an artifact. This is what I adopted—simple implementation and no dependency on the script.

Align with script end position. Derive the script end position from Whisper segments and check if energy exists beyond that point. This is more general but reintroduces STT dependency. If artifacts don’t appear in Whisper segments, the end position might be incorrectly determined.

After removing false positives:

Before (with commas): 29 / 72 detections   ← all 12 models triggered
After (single sentences only): 16 / 72 detections   ← only 4 models triggered

Had I reported the initial results as-is, I would have spread false panic of “all 12 failed.” Once you build a detector, you must first test it on things that should not trigger it.

What I learned is that audio quality inspection requires multiple methods that reveal different layers:

Relying only on STT for QC meant assuming everything visible in the most familiar tool would be visible everywhere. In reality, STT only shows “what can be recognized as words.”

Interestingly, these 4-second artifacts are hard to notice even when listening. A 0.1-second sound feels like “some lingering resonance” unless you’re paying close attention. Only by laying out the numbers do you realize there’s an abnormal structure: “silence 0.5s followed by sound.”

If humans can’t perceive it subjectively, machines must measure it. And every measurement method has its own blind spots.

Reflecting on this experience, here’s the process I should have followed:

The fourth point is the key lesson: when detection rates are too high, suspect the detector, not the targets.

A record of designing voices from single captions, manufacturing training corpora, and mass-producing role-specific practical voices. This article is Part 3: Quality Gate.

← Previous: Weeding out candidates using fixable defects

→ Next: How 70 minutes of training material vanished in an instant due to a network blink

Full series (18 parts)

The insights in this article are compiled in the Diffusion TTS Manufacturing Pipeline for Mass-producing Practical Voices.