You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Chatterbox Multilingual MeanFlow Streaming (experimental)

Private experimental package combining Chatterbox's Spanish-specialized T3 text-to-speech token generator with Chatterbox Flash's distilled two-step MeanFlow S3Gen decoder. The package is configured for warm, low-latency inference on an NVIDIA L4.

Tested setting

First decode at four speech tokens (40 ms playable audio).
Two MeanFlow steps.
Three look-ahead tokens, one duplicated boundary token, 40 ms crossfade.
Preferred two-step configuration measured 172.21 ms p95 to its first 40 ms of playable audio in the controlled L4 test.
The full 4.4-second streamed decode ran about 7.0 times faster than playback.

This is experimental, single-request, warm-GPU evidence. Network transport, telephony buffers, cold start, concurrency, WER, speaker similarity, and formal listening quality are not yet validated.

Before/after audio

The voice in every player below is the selected Mateo project voice: CML-TTS speaker 12367, using the user-selected ref_04_cml_8860.wav. All three samples use this same conditioning audio, Spanish text, and exact T3 speech-token sequence, so the comparison isolates the decoder change as closely as possible.

Test sentence: Su reserva está confirmada para el viernes tres a las ocho y media de la tarde. Le enviaremos un recordatorio al móvil el día anterior.

In blind listening, the user described these three preferred-token samples as barely distinguishable. An earlier disliked draw sounded equally poor through the original, full MeanFlow, and streamed decoders; this indicates that T3 sampling/prosody consistency is the remaining quality risk, not the stream joins.

For listening criteria and raw amplitude measurements, see the audio_comparison guide.

Installation

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python apply_callback_patch.py

Test

Use a clean reference WAV containing one speaker:

python inference_streaming.py \
  --voice reference_voice/ref_04_cml_8860.wav \
  --language es \
  --text "Hola, le llamo para confirmar su cita." \
  --output test.wav

For live calling, pass clause-sized text to TTS as it becomes available. Supplying an entire paragraph increases T3 prefill and can push first audio above 200 ms.

Reference voice provenance

The included Mateo reference is an excerpt associated with CML-TTS speaker 12367, selected by the user from the Mateo_TTS project's Castilian-Spanish references. CML-TTS is LibriVox-derived and is distributed under CC BY 4.0; this excerpt has been selected and processed for voice conditioning. Attribution: CML-TTS contributors and the source LibriVox recording. Do not use this package to identify the original narrator.

Files

t3_es_es.safetensors: Spanish-specialized T3 weights.
s3gen_meanflow.safetensors: distilled Flash MeanFlow decoder weights.
ve.pt: voice encoder.
grapheme_mtl_merged_expanded_v1.json: multilingual tokenizer vocabulary.
config.json: tested streaming and sampling settings.
inference_streaming.py: integrated callback, streaming decode, and crossfade.
apply_callback_patch.py: minimal callback installer for the pinned source revision.

Based on Resemble AI's Chatterbox and Chatterbox Flash repositories. Preserve and review upstream licensing and model-card requirements before redistribution or production use.

Downloads last month: -