cstr/qwen3-tts-0.6b-base-GGUF

Qwen3-TTS 12Hz 0.6B Base — GGUF (CrispASR)

GGUF / ggml conversions of `Qwen/Qwen3-TTS-12Hz-0.6B-Base` for use with the qwen3-tts backend in [CrispStrobe/CrispASR](https://github.com/CrispStrobe/CrispASR).

Qwen3-TTS 12Hz 0.6B Base is Qwen's multilingual voice-cloning TTS model:

• —10 supported languages: zh en ja ko de fr ru pt es it
• —discrete multi-codebook LM architecture with a separate 12 Hz tokenizer / codec
• —runtime voice cloning from (ref_audio, ref_text) or pre-baked voice-pack GGUFs
• —Apache-2.0 licence

This repo contains the talker / code-predictor / speaker-encoder model. It must be used together with the separate tokenizer / codec GGUF from `cstr/qwen3-tts-tokenizer-12hz-GGUF`.

Files

Quick Start

Build CrispASR:

git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc) --target crispasr

Download the talker + tokenizer:

huggingface-cli download cstr/qwen3-tts-0.6b-base-GGUF \
    qwen3-tts-12hz-0.6b-base-q8_0.gguf --local-dir .

huggingface-cli download cstr/qwen3-tts-tokenizer-12hz-GGUF \
    qwen3-tts-tokenizer-12hz.gguf --local-dir .

Voice clone from a reference WAV:

./build/bin/crispasr \
    --backend qwen3-tts \
    -m qwen3-tts-12hz-0.6b-base-q8_0.gguf \
    --codec-model qwen3-tts-tokenizer-12hz.gguf \
    --voice clone.wav \
    --ref-text "Exact transcript of clone.wav" \
    --tts "Hello there" \
    --tts-output hello.wav

Use a baked voice-pack GGUF:

./build/bin/crispasr \
    --backend qwen3-tts \
    -m qwen3-tts-12hz-0.6b-base-q8_0.gguf \
    --codec-model qwen3-tts-tokenizer-12hz.gguf \
    --voice my-voice-pack.gguf \
    --tts "Hello there" \
    --tts-output hello.wav

When --voice points to a .wav, --ref-text is required. When --voice points to a .gguf, it is treated as a baked voice pack.

Quantisation Notes

Current CrispASR validation status:

• —qwen3-tts-12hz-0.6b-base.gguf
• —reference baseline
• —qwen3-tts-12hz-0.6b-base-q8_0.gguf
• —recommended quantised deployment
• —end-to-end synthesis is audibly good in current CrispASR testing
• —intermediate activations still drift measurably from F16 in strict tensor diffs
• —qwen3-tts-12hz-0.6b-base-q4_k.gguf
• —smallest variant; loads end-to-end
• —content fidelity is significantly worse than Q8_0. In a fixed- seed back-to-back A/B (JFK voice prompt, "Hello world, this is a quick speed benchmark for the qwen three TTS pipeline.", 8 s output), the F16 talker rendered the first ~5 words correctly, Q80 rendered the first ~4 words, and Q4K produced audio that a strong ASR (Qwen3-ASR-0.6B) could not classify as English. Q4K may recover when given more frames (longer warmup) or longer prompts, but for short utterances it is **not** a drop-in replacement for Q80.
• —not numerically faithful to F16 in strict tensor diffs.
• —other lower-bit talker quants (q6_k, q5_k) are similarly experimental and not currently shipped — convert with crispasr-quantize if you need them.

If fidelity matters more than memory:

• —use qwen3-tts-12hz-0.6b-base.gguf
• —keep the companion tokenizer / codec at F16

If you want the best currently-tested size / quality trade-off:

• —use qwen3-tts-12hz-0.6b-base-q8_0.gguf
• —still keep qwen3-tts-tokenizer-12hz.gguf at F16

If disk space is the binding constraint:

• —use qwen3-tts-12hz-0.6b-base-q4_k.gguf
• —still keep qwen3-tts-tokenizer-12hz.gguf at F16 (quantising the codec hurts earlier than quantising the talker — runtime_ref_codes is sensitive)

How this was made

1. 1.The upstream HF safetensors checkpoint was converted to GGUF F16 by `models/convert-qwen3-tts-to-gguf.py`.
2. 2.Quantised variants are produced with CrispASR's GGUF quantiser.
3. 3.Inference is implemented in `src/qwen3_tts.cpp`, using ggml graphs for the talker / code-predictor path and the companion tokenizer GGUF for codec encode/decode.

Reference implementation

Architecture and behaviour were checked against the official Qwen release:

• —upstream model card: `Qwen/Qwen3-TTS-12Hz-0.6B-Base`
• —upstream repository: `QwenLM/Qwen3-TTS`
• —technical report: `Qwen3-TTS Technical Report`

The CrispASR runtime is a clean C++ / ggml re-implementation for this repo's backend stack.

Related

• —Companion tokenizer / codec GGUF: `cstr/qwen3-tts-tokenizer-12hz-GGUF`
• —Upstream Base model: `Qwen/Qwen3-TTS-12Hz-0.6B-Base`
• —C++ runtime: `CrispStrobe/CrispASR`

License

Apache-2.0, inherited from the base model.

Provenance and EU AI Act Art. 53 note

• —Upstream model: Qwen/Qwen3-TTS-12Hz-0.6B-Base — published by Qwen.
• —Upstream licence: apache-2.0. 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.
• —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.