Kokoro: short prompts (~5-8 words) produce run-on garbage after the text

[thedecentthing]

Summary

Kokoro generations with short prompts (≈5–8 words) produce audio that continues past the text: the sentence is spoken correctly, then repeated, then followed by multilingual noise. Longer prompts (≈10+ words) on the same voice are clean. Reproduces on two independent compute paths (the app's MLX bundle and a from-source backend running the kokoro pip package on pytorch CPU), so it looks like a pipeline/model-behavior issue rather than an acceleration-path one.

Environment

• macOS, Apple Silicon
• Voicebox.app 0.5.0 (MLX backend) — also reproduced on a from-source backend (kokoro>=0.9.4, pytorch CPU) — same ~2.9 s file size, same garbage content
• Model: hexgrad/Kokoro-82M (314 MB, HF_HUB_DISABLE_XET=1 prefetch; also reproduced with in-app download)
• Voices tested: af_sarah, af_bella presets

Steps to reproduce

curl -X POST http://127.0.0.1:17493/speak \
  -H 'content-type: application/json' \
  -d '{"text": "Read it back to me. Word for word.", "profile": "<kokoro-preset-profile-id>", "engine": "kokoro"}'
# fetch the WAV at /audio/{id} when complete, then transcribe

Actual result

Transcript of the returned audio (sentence repeated, then noise):

Read it back to me, word for word… Read it back to me, word for word.! Out! Out! End- OB 7

RMS energy profile (100 ms windows) — the audio continues with real energy well past the text:

____.##.#####._.##.###._______
^text        ^garbage  ^end

Control with a longer prompt on the same voice ("We never authorized that deployment. The model card is explicit, and your team knew it.") is perfectly clean, 5.97 s, no tail.

Additional observations

• Both takes of a seeded bake (seed: 0 and seed: 1) show the same behavior.
• The garbage tail appears regardless of whether the request goes through /speak directly or through a full client pipeline, so the client is not implicated.
• Other short lines (5–8 words) show the same pattern; the boundary in our tests is roughly ≤8 words.
• Whisper STT on these files faithfully reports the extra audio, so any transcript-diff tooling downstream flags the take.

Question

Is this a known Kokoro-82M behavior on short prompts (EOS/stop not emitted), and are there generation knobs Voicebox should be setting for short text (max frames/tokens, stop conditions, minimum-length padding)? Happy to test patches against this machine.