chatterbox-nano-GGUF

Chatterbox-Nano TTS — GGUF (ggml)

GGUF / ggml conversion of ResembleAI/chatterbox-nano for use with CrispStrobe/CrispASR (v0.8.31+, issue #382).

Chatterbox-Nano is the small sibling of Chatterbox-Turbo: the same AR text-to-speech pipeline with a GPT2-small T3 backbone (12 layers, 768 hidden, 12 heads) instead of Turbo's GPT2-medium (24L, 1024D). Upstream's s3gen_meanflow.safetensors, ve.safetensors and conds.pt are byte-identical to ResembleAI/chatterbox-turbo (same LFS blobs), so this repo ships only the T3 — use the S3Gen from cstr/chatterbox-turbo-GGUF as the companion. CrispASR's registry does this automatically. Distributed under MIT license.

Files

The GGUF carries chatterbox.t3.n_kv_heads = 12 explicitly (GPT-2 is MHA; CrispASR's built-in default of 16 fits Turbo only). Token ids match upstream inference (T3Config: start_text 255, stop_text 0, speech 6561/6562) — the stop_text_token: 50256 in t3_nano_v1.yaml is a training-time artifact that upstream generate() never reads.

Usage

# auto-download T3 + Turbo S3Gen companion
crispasr --backend chatterbox-nano -m auto --auto-download \
  --tts "Hello from Chatterbox Nano." --tts-output nano.wav

# explicit files
crispasr --backend chatterbox-nano -m chatterbox-nano-t3-q8_0.gguf \
  --codec-model chatterbox-turbo-s3gen-q8_0.gguf \
  --tts "Hello from Chatterbox Nano." --tts-output nano.wav

Voice cloning (--voice reference.wav), the 19 [emotion] style tokens, and the sampler defaults are identical to the Turbo backend; CFG/min_p/exaggeration are not supported by the Turbo/Nano pipeline (upstream disables them).

Conversion

Converted with models/convert-chatterbox-to-gguf.py --variant nano from upstream revision 71ccd1d0; quantized with crispasr-quantize.