Qwen3-TTS-12Hz-0.6B-CustomVoice — GGUF (ggml-quantised)
The CustomVoice variant is the fixed-speaker sibling of Qwen3-TTS-12Hz-0.6B-Base: instead of cloning a voice from a 3 s reference WAV (the Base path), it ships 9 baked speaker tokens picked via a --voice <name> flag. No ECAPA forward, no codec encoder, no reference audio required. Two of the speakers (dylan, eric) carry Chinese-dialect overrides (Beijing / Sichuan) that re-route language_id when synthesising Chinese-or-auto.
| Speaker | Language / dialect |
|---|
aiden (default) | English (M) |
dylan | Beijing dialect (M, dialect_token=2074) |
eric | Sichuan dialect (M, dialect_token=2062) |
ono_anna | English (F) |
ryan | English (M) |
serena | English (F) |
sohee | English (F) |
uncle_fu | English (M, older) |
vivian | English (F) |
Pair this with the codec at
cstr/qwen3-tts-tokenizer-12hz-GGUF — the talker emits 16-codebook RVQ codes that the codec decoder renders to 24 kHz PCM.
Files
| File | Quant | Size | Notes |
|---|
qwen3-tts-12hz-0.6b-customvoice-q8_0.gguf | Q8_0 | 968 MB | Recommended — ASR-roundtrip word-exact vs reference |
The 1.7B-CustomVoice variant ships at
cstr/qwen3-tts-1.7b-customvoice-GGUF. For ICL voice cloning use
cstr/qwen3-tts-0.6b-base-GGUF instead.
Quick start
1# 1. Build CrispASR
2git clone https://github.com/CrispStrobe/CrispASR
3cd CrispASR
4cmake -B build -DCMAKE_BUILD_TYPE=Release
5cmake --build build -j --target crispasr
6
7# 2. Pull the talker + the codec
8huggingface-cli download cstr/qwen3-tts-0.6b-customvoice-GGUF qwen3-tts-12hz-0.6b-customvoice-q8_0.gguf --local-dir .
9huggingface-cli download cstr/qwen3-tts-tokenizer-12hz-GGUF qwen3-tts-tokenizer-12hz.gguf --local-dir .
10
11# 3. Synthesise — pick a speaker by name
12./build/bin/crispasr --backend qwen3-tts-customvoice \
13 -m qwen3-tts-12hz-0.6b-customvoice-q8_0.gguf \
14 --codec-model qwen3-tts-tokenizer-12hz.gguf \
15 --voice ryan \
16 --tts "Hello, this is the Ryan speaker." \
17 --tts-output ryan.wav
For auto-download simply pass -m auto:
1./build/bin/crispasr --backend qwen3-tts-customvoice -m auto \
2 --voice serena \
3 --tts "Auto-download fetches both files." \
4 --tts-output out.wav
Quality verification
| Speaker | Synthesised text | Parakeet output |
|---|
vivian | "Hello, this is a CustomVoice test using the vivian speaker." | "Hello! This is a custom voice test using the Vivian speaker." (verbatim modulo case/punct) |
aiden | "The quick brown fox jumps over the lazy dog." | "The quick brown fox jumps over the lazy dog." (verbatim) |
serena | "Testing the new backend alias and the serena speaker." | "Testing the new back end Ilias and the Serena speaker." (1 ASR misrecognition; audio clean) |
dylan | "你好，今天天气真不错。" | dialect override engaged (language_id=2074); 3.28 s clean audio |
Architecture
| Component | Details |
|---|
| Talker LM | Qwen3 (28 layers, 1024 hidden, 16 heads, 8 KV heads, head_dim=64) |
| Output head | 16 codebooks × 2048 (RVQ) — emits codes for the codec |
| Code predictor | 5L Qwen3 + 15 separate codec_embedding/lm_head pairs (top-k=50, temp=0.9 sampling) |
| Codec | Qwen3-TTS-Tokenizer-12Hz (separate GGUF, 12.5 fps RVQ) |
| Audio | 24 kHz mono float32 PCM |
CustomVoice differs from Base in that the speaker embedding is not computed via ECAPA on a reference WAV. Instead talker.get_input_embeddings()(spk_id) retrieves a row from the codec embedding table directly (modeling_qwen3_tts.py:2091). No reference audio, no codec encoder needed at runtime.
CrispASR backend integration
| Flag | Purpose |
|---|
--backend qwen3-tts-customvoice | Selects the CustomVoice runtime path |
--voice <name> | Picks a fixed speaker by name (default: first speaker, aiden) |
--tts "..." | Text to synthesise |
--codec-model PATH | Qwen3-TTS-Tokenizer-12Hz GGUF |
--tts-output PATH | Output WAV (24 kHz mono) |
The runtime branches on qwen3tts.tts_model_type ("custom_voice" vs "base" vs "voice_design") so the same backend object handles all three Qwen3-TTS variants.
Conversion
1python models/convert-qwen3-tts-to-gguf.py \
2 --input Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice \
3 --output qwen3-tts-12hz-0.6b-customvoice-f16.gguf \
4 --outtype f16
5
6build/bin/crispasr-quantize qwen3-tts-12hz-0.6b-customvoice-f16.gguf \
7 qwen3-tts-12hz-0.6b-customvoice-q8_0.gguf q8_0
The converter writes qwen3tts.tts_model_type=custom_voice plus qwen3tts.spk_names, qwen3tts.spk_token_ids, and qwen3tts.spk_dialect_token_ids so the runtime can resolve --voice <name> to the right speaker token + dialect override at load time.
Attribution
- Talker base:
Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice (Apache 2.0).
- Codec base:
Qwen/Qwen3-TTS-Tokenizer-12Hz (Apache 2.0) — see cstr/qwen3-tts-tokenizer-12hz-GGUF.
- Reference engine: QwenLM/Qwen3-TTS (
modeling_qwen3_tts.py).
- GGUF conversion + ggml runtime:
CrispStrobe/CrispASR — see src/qwen3_tts.cpp, models/convert-qwen3-tts-to-gguf.py.
License
Apache 2.0 — both the upstream talker and codec, and the GGUF conversion.
Provenance and EU AI Act Art. 53 note
- Upstream model: Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice — 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. 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.
