Voxtral-4B-TTS-2603 — GGUF
GGUF / ggml conversions of mistralai/Voxtral-4B-TTS-2603 for use with the voxtral-tts backend of CrispStrobe/CrispASR — one C++ binary, no Python.
Text-to-speech across 9 languages (en, fr, de, es, it, pt, nl, ar, hi) with 20 preset voices, 24 kHz output.
Architecture
A three-component pipeline, all implemented as ggml compute graphs:
- LLM backbone — Ministral-3B autoregressive decoder (26 layers, GQA 32/8, NORMAL/adjacent-pair RoPE θ=1e6). Conditioned on a preset voice prefix + text, it emits one hidden state per audio frame.
- Flow-matching acoustic transformer — 3-layer bidirectional transformer (no positional encoding). Per frame: a semantic token (greedy argmax) plus 36 acoustic FSQ codes via an 8-step (7-interval) Euler ODE with classifier-free guidance (α=1.2).
- Voxtral codec decoder — 292-d input (256-d semantic VQ + 36-d FSQ) → causal conv → 4× [2-layer ALiBi transformer + ConvTranspose1d upsampling] → 240 PCM samples/frame at 24 kHz.
Files
| File | Precision | Size | Notes |
|---|---|---|---|
voxtral-4b-tts-q4_k.gguf | Q4_K | ~2.4 GB | default; fits comfortably in 8 GB RAM |
voxtral-4b-tts-q8_0.gguf | Q8_0 | ~4.3 GB | higher quality |
voxtral-4b-tts-f16.gguf | F16 | ~8.2 GB | reference precision |
The semantic VQ codebook (codec.semantic_cb.weight) and preset voice embeddings are kept at F32 in every file.
Usage
# auto-downloads the Q4_K on first run
crispasr --backend voxtral-tts -m auto --auto-download \
--tts "Bonjour le monde." --voice fr_female --tts-output out.wav
--voice accepts any preset (e.g. neutral_female, neutral_male, casual_female, cheerful_female, fr_female, de_male, es_female, it_male, pt_female, nl_male, ar_male, hi_female, …). Omit it for the default voice. Add --seed N for a reproducible acoustic sample.
License
The GGUF conversions inherit the base model's CC-BY-NC-4.0 license (non-commercial). See mistralai/Voxtral-4B-TTS-2603 for the original weights and terms.
Provenance and EU AI Act Art. 53 note
- Upstream model: mistralai/Voxtral-4B-TTS-2603 — published by
mistralai. - Upstream licence:
cc-by-nc-4.0. 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.