Modelpublic

cstr/qwen3-tts-1.7b-customvoice-GGUF

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

7likes2.7kdownloads

Model Card

Qwen3-TTS-12Hz-1.7B-CustomVoice — GGUF (ggml-quantised)

GGUF / ggml conversion of `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` for use with [CrispStrobe/CrispASR](https://github.com/CrispStrobe/CrispASR).

The 1.7B sibling of `cstr/qwen3-tts-0.6b-customvoice-GGUF`. Same fixed-speaker pattern (9 baked speakers, no ECAPA, no reference WAV), but on the 1.7B talker — hidden=2048, head_dim=128, with a small_to_mtp_projection MLP that maps the 2048-d talker hidden state down to the 1024-d code predictor input space.

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-1.7b-customvoice-f16.gguf`	F16	3.84 GB	Reference quality
`qwen3-tts-12hz-1.7b-customvoice-q8_0.gguf`	Q8_0	2.04 GB	Recommended — ASR-roundtrip word-exact vs F16

The 0.6B-CustomVoice variant ships at `cstr/qwen3-tts-0.6b-customvoice-GGUF`. The 1.7B-Base (ICL voice clone) variant ships at `cstr/qwen3-tts-1.7b-base-GGUF`.

Quick start

bash

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

# 2. Pull the talker + the codec
huggingface-cli download cstr/qwen3-tts-1.7b-customvoice-GGUF qwen3-tts-12hz-1.7b-customvoice-q8_0.gguf --local-dir .
huggingface-cli download cstr/qwen3-tts-tokenizer-12hz-GGUF qwen3-tts-tokenizer-12hz.gguf --local-dir .

# 3. Synthesise — pick a speaker by name
./build/bin/crispasr --backend qwen3-tts-1.7b-customvoice \
    -m qwen3-tts-12hz-1.7b-customvoice-q8_0.gguf \
    --codec-model qwen3-tts-tokenizer-12hz.gguf \
    --voice ryan \
    --tts "Hello, this is the Ryan speaker on the 1.7B model." \
    --tts-output ryan.wav

For auto-download simply pass -m auto:

bash

./build/bin/crispasr --backend qwen3-tts-1.7b-customvoice -m auto \
    --voice vivian \
    --tts "Auto-download fetches both files." \
    --tts-output out.wav

Quality verification

ASR roundtrip via `cstr/parakeet-tdt-0.6b-v3-GGUF`:

Variant	Speaker	Output (verbatim)
Q8_0	`ryan`	`"Hello, this is the Ryan speaker."`
F16	`vivian`	`"This is a longer prompt to verify the 1.7B CustomVoice path through the talker."`

Both roundtrips are exact-match on a long prompt (the regression test for the small_to_mtp_projection per-step fix in commit 2cc7aeb).

Architecture

Component	Details
Talker LM	Qwen3 (28 layers, 2048 hidden, 16 heads, 8 KV heads, head_dim=128)
Output head	16 codebooks × 2048 (RVQ) — emits codes for the codec
`small_to_mtp_projection`	2048-d → 1024-d MLP (input bridge to code predictor) — applied per-step at decode (steps 1..14, fix in `2cc7aeb`)
Code predictor	5L Qwen3 (1024 hidden) + 15 separate codecembedding/lmhead pairs (top-k=50, temp=0.9)
Codec	Qwen3-TTS-Tokenizer-12Hz (separate GGUF, 12.5 fps RVQ)
Audio	24 kHz mono float32 PCM

CustomVoice 1.7B uses the same fixed-speaker contract as 0.6B-CV: talker.get_input_embeddings()(spk_id) retrieves a row from the codec embedding table directly. The 2048-d hidden state then gets projected down to 1024-d via small_to_mtp_projection before each code predictor step.

CrispASR backend integration

Flag	Purpose
`--backend qwen3-tts-1.7b-customvoice`	Selects the 1.7B 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 and on the GGUF talker hiddendim (`spkenc_dim`) so the same backend object handles 0.6B/1.7B Base/CustomVoice/VoiceDesign automatically.

Conversion

bash

python models/convert-qwen3-tts-to-gguf.py \
    --input Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
    --output qwen3-tts-12hz-1.7b-customvoice-f16.gguf \
    --outtype f16

build/bin/crispasr-quantize qwen3-tts-12hz-1.7b-customvoice-f16.gguf \
    qwen3-tts-12hz-1.7b-customvoice-q8_0.gguf q8_0

Attribution

—Talker base: `Qwen/Qwen3-TTS-12Hz-1.7B-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-1.7B-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.