BTimber/piper-en_US-mira

Mira (en_US) — a Piper voice that actually laughs

A free Piper voice (VITS, medium, 22.05 kHz, LJSpeech-derived timbre) with trained non-verbal expressions: real laughs, sighs, gasps, groans, and throat-clears — produced by the vocal model itself, not spliced-in sound effects.

🔊 Listen: samples/laugh_demo.wav · samples/sigh_throat_demo.wav

What it does that other Piper voices don't

Standard Piper voices can only read text. This voice was fine-tuned in phoneme_ids mode with unused phoneme IDs repurposed as non-verbal marker tokens, then trained on an emotive corpus (distilled from ChatterBox Turbo, MIT) in which those markers align with genuine non-verbal audio. Inject a marker at synthesis time and the model performs it, in-voice, with natural prosody flowing through it:

Tag	Phoneme id	Marker char	Repeat
[laugh]	39	æ	5
[sigh]	40	ç	3
[gasp]	42	ø	2
[groan]	43	ħ	3
[clear throat]	45	œ	3

(markers.json in this repo is the canonical map.)

Usage

Plain speech works like any Piper voice:

echo "The quick brown fox jumps over the lazy dog." | \
  piper -m en_US-mira-tokens2.onnx -f out.wav

Non-verbals need id-level synthesis — normal Piper phonemizes text through espeak, which can never emit the marker characters. mira_synth.py (MIT, included) is the reference implementation: it phonemizes the text segments, splices the marker char (× its repeat count) where a [tag] appears, converts to ids once, and calls phoneme_ids_to_audio:

from piper import PiperVoice
import json
from mira_synth import synth_expressive

voice = PiperVoice.load("en_US-mira-tokens2.onnx")
markers = json.load(open("markers.json"))
wav = synth_expressive(voice, markers, "Oh that's funny [laugh] tell me the rest.")
open("out.wav", "wb").write(wav)

⚠️ The phoneme-dialect trap (read this before fine-tuning or re-exporting)

This model's phoneme inventory was built with espeak voice en — NOT en-us. The shipped config already says "voice": "en". Do not change it, and do not phonemize inputs with en-us. Two things break if you do:

1. en-us emits æ for TRAP vowels ("half", "crash", "laugh"…) — and in this voice æ is repurposed as the laugh token (id 39). Phonemize with the wrong dialect and TRAP-vowel words trigger giggles mid-word or collapse into mangled forms ("half" → "huff"-like artifacts). This is not hypothetical; we shipped that bug to ourselves once.
2. en-us adds diphthong symbols (aɪ aʊ ɔɪ eɪ oʊ) that map to ids this model never trained — silent quality damage on any word containing them.

If you fine-tune further: pass --data.espeak_voice en to the Piper trainer, and note that some trainer versions regenerate the exported config from the training flag — if your flag says en-us, the exported .onnx.json will too, and the corruption becomes the model's default. Check the exported config's "voice" field before deploying. The extra en-us symbol entries a bad export adds (diphthongs at ids 161–165) are the fingerprint of this mistake.

Lineage & licensing

• Base: Piper en_US-ljspeech-medium checkpoint — LJSpeech dataset is public domain.
• Emotive teacher: ChatterBox Turbo (MIT) — ~520-clip distillation corpus conditioned on the LJSpeech timbre, including genuine non-verbal vocalizations aligned to the marker tokens.
• Training: Piper trainer, phoneme_ids mode, marker tokens as above.
• Weights: CC0-1.0 (public domain dedication). mira_synth.py: MIT.
• The Piper runtime is a separate install and carries its own license.

Made in a home lab as part of an embodied-companion project; shared so other Piper users can have voices that breathe. No cloud, no strings.