Chatterbox-Turbo TTS — GGUF (ggml)

GGUF / ggml conversion of ResembleAI/chatterbox-turbo for use with CrispStrobe/CrispASR.

Chatterbox-Turbo is a distilled 350M-parameter TTS pipeline: GPT-2 tokenizer + AR text-to-speech model + meanflow S3Gen (2-step CFM, vs 10 for base Chatterbox) + HiFTGenerator vocoder. Distributed under MIT license.

Two GGUF files are needed: the T3 model (text to speech tokens) and the S3Gen model (speech tokens to audio).

Updates

2026-06-21 — tokenizer fix (re-uploaded T3 files). The T3 GGUFs previously embedded only the 50257-token base GPT-2 vocab, while the T3 text embedding is 50276 — the 19 extra ids are the turbo emotion/style control tokens. The files have been re-uploaded with the full 50276-token tokenizer (base vocab + added_tokens.json), so they are now internally consistent and load cleanly on strict loaders (CrispASR ≥ v0.8.1). The weights are unchanged (byte-for-byte), so this is a tokenizer-only update.

Emotion / style tags. You can drive prosody by putting any of these bracketed tags in the input text (CrispASR ≥ v0.8.1 emits them as their special token id):

[laugh] [chuckle] [sigh] [gasp] [cough] [groan] [sniff] [shush] [clear throat]
[whispering] [angry] [happy] [crying] [fear] [surprised] [sarcastic] [dramatic]
[narration] [advertisement]

Example: [laugh] Thank you so much! prepends a laugh to the line. (Effect strength varies per tag and prompt.)

Files

Encoder attention/FFN weights are stored at F32 precision for quality. Vocoder weights (conv_pre, resblocks, conv_post, source fusion, F0 predictor) are F32.

Quick start

# 1. Build CrispASR
git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF
cmake --build build -j --target chatterbox

# 2. Pull both model files (Q8_0 recommended)
huggingface-cli download cstr/chatterbox-turbo-GGUF chatterbox-turbo-t3-q8_0.gguf --local-dir .
huggingface-cli download cstr/chatterbox-turbo-GGUF chatterbox-turbo-s3gen-q8_0.gguf --local-dir .

# 3. Synthesise (C API — CLI adapter in progress)
# See test programs in SESSION_HANDOVER.md for usage examples

Architecture

Text -> GPT-2 BPE tokenizer (50276 tokens: 50257 base + 19 emotion/style tags)
     -> T3 GPT-2 AR (24 layers, 1024D, 16 heads, learned pos emb, SwiGLU)
     -> 25 Hz speech tokens (6561 codebook)
     -> UpsampleConformerEncoder (6 pre + 4 post upsample, 512D, 8 heads, rel-pos attn)
        -> Upsample1D: nearest-neighbor 2x + Conv1d(512,512,k=5) + Linear + LayerNorm + xscale
     -> 80-channel mel spectrogram (50 Hz)
     -> Meanflow CFM denoiser (2 Euler steps, linear schedule, no CFG)
        UNet1D: 1 down + 12 mid + 1 up blocks, 256 ch, 4 transformer blocks each
     -> HiFTGenerator vocoder (F0 predictor + SineGen + 3x ConvTranspose1d + iSTFT)
     -> 24 kHz mono WAV

Key differences from base Chatterbox

Quality verification

ASR roundtrip using same speech tokens as Python reference:

Conversion

# From HuggingFace model (requires chatterbox-tts pip package):
python models/convert-chatterbox-to-gguf.py \
  --input ResembleAI/chatterbox-turbo \
  --output-dir /path/to/output \
  --variant turbo

Related models

• cstr/chatterbox-GGUF — base Chatterbox (Llama T3, 10-step CFM)
• cstr/lahgtna-chatterbox-v1-GGUF — Arabic T3 variant

Provenance and EU AI Act Art. 53 note

• Upstream model: ResembleAI/chatterbox-turbo — published by ResembleAI.
• Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.