AIToday
Audio & SpeechQiita 機械学習Published: Nov 22, 2026, 22:01 JST

Irodori-TTS loses 4–7dB above 4kHz; fixing the reference audio, not settings, works

Send on LINE
Irodori-TTS loses 4–7dB above 4kHz; fixing the reference audio, not settings, works

3 Key Points

  1. What happened

    Testing Irodori-TTS, the author found it speaks 4–7dB quieter above 4kHz than its reference (―31.9dB vs ―22.9dB), even though settings like the Echo-TTS "crisp" preset (--rescale-k 0.96), more steps, or 64-step guidance moved brightness by less than 1dB.

  2. Why it matters

    The problem lies in the model's own behavior, which users cannot change directly; but pre-processing the reference audio to compensate for the loss made the output match the original recording's brightness, offering a practical workaround for anyone using this voice-cloning tool.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The author, a non-engineer working with AI assistance, set out to fix two issues with Irodori-TTS: a muffled sound and a lack of emotional expression. By measuring the average tone of the reference and the model's output, they pinpointed that the model loses 4–7dB above 4kHz, with only 2–3dB attributable to the audio encoding step; the rest occurs in the audio generation. Common settings adjustments, including those suggested for the underlying Echo-TTS model, failed to improve brightness. Instead, pre-boosting the reference audio's high frequencies compensated for the loss, and the output matched the original recording's brightness. For emotional expression, the author found that using a reference generated by the model itself, rather than a real recording, allowed for acting. This fix introduced additional muffling, which was resolved by boosting only frequencies above 4kHz to avoid distortion in the 2–4kHz range. The author notes limitations: only two voices (both their own) and three test sentences were used, and the results may not generalize. The approach offers a practical workaround for similar voice cloning tools when model internals cannot be changed.

FAQ
What settings did not fix the muffled sound?
Settings like --rescale-k 0.96, --truncation-factor 0.9, increasing --cfg-scale-speaker to 7.0, using --cfg-min-t 0.2, and --num-steps 64 all changed brightness by less than 1dB. Even a text prompt asking for clear audio had little effect.
How was the muffled sound fixed?
The author measured the difference in average tone between the reference and the model's spoken output, then boosted the reference audio's high frequencies by that amount before feeding it to the model. This made the output match the original recording's brightness (e.g., above 8kHz went from −36.5dB to −29.5dB).
Why did the voice lack emotional expression?
The model struggles to act when the reference is from a real recording. Replacing the reference with audio the model itself generated (reading the same sentences) allowed for more expressive output, though it introduced muffling that was fixed by boosting only frequencies above 4kHz.
Qiita 機械学習Read Original Article
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleWells Fargo tells customers: AI agent use is AT YOUR OWN RISK