Tts trailing silence and hallucination trim

Skill kjuhwa/skills-hub/skills/ml-ops/tts-trailing-silence-and-hallucination-trim

Self-correcting knowledge corpus for Claude Code — 9 stable shape clusters, bias-correction pipeline baked into contribution flow. 47 papers, 45 techniques, 1.1k skills.

Install

npx -y skills add kjuhwa/skills-hub --skill tts-trailing-silence-and-hallucination-trim

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Post-process TTS output to cut long internal silences plus any trailing hallucinated noise with a cosine fade-out.

SKILL.md

2.8 KB, as published. Nobody here has run it

TTS trailing silence and hallucination trim

When to use

Your TTS engine (Chatterbox, Chatterbox Turbo, some variants of XTTS) occasionally produces [speech][silence][hallucinated noise]: a long pause followed by unrelated mumbling, breathing, or music. You want a deterministic, no-ML cleanup pass that cuts the output at the natural speech boundary and fades out cleanly.

Steps

Compute per-frame RMS energy over a short window (e.g. frame_ms=20, so frame_len = sample_rate * frame_ms / 1000). Convert a silence_threshold_db (e.g. -40 dB) to linear amplitude and classify each frame as speech or silence.
Find the first speech frame and step back by one frame for padding — that becomes the new start sample. This handles small leading silences without cutting the first phoneme.
Walk forward from the first speech frame, counting consecutive silence frames. The first time consecutive_silence >= max_internal_silence_frames (e.g. 1000 ms), set the cut point at the start of that long silence and stop. This is the key insight: hallucinations are gated by a long gap.
From the cut point, trim back any trailing silence until you find a speech frame, then add a short tail (e.g. 200 ms of min_silence_ms) so the ending doesn't sound truncated.
Apply a short cosine fade-out (~30 ms) on the last N samples: fade = cos(linspace(0, pi/2, N)) ** 2 multiplied into the tail. The cosine-squared curve is inaudible; a linear fade clicks.
Expose the trim as a pure function (audio, sample_rate) -> audio so it can be plugged into a chunked-TTS wrapper or applied to a whole generation.

Counter / Caveats

Only enable this for engines that actually hallucinate (flag them with a needs_trim=True in your model registry). Clean engines will lose legitimate dramatic pauses.
The threshold-based detector cannot tell "hallucinated noise" from "real faint speech". A human with a whisper/breath pattern may get cut. If users report truncation, relax silence_threshold_db first, then max_internal_silence_ms.
Use np.sqrt(mean(audio**2)) per frame, not instantaneous peak — peaks over-react to transients.
Keep the fade-out short (~30 ms). Longer fades chew into the final syllable.

Source references: backend/utils/audio.py (trim_tts_output), backend/utils/chunked_tts.py (uses trim_fn per chunk).