qwen3tts — Qwen3-TTS text-to-speech from GGUF¶

The Zig port of qwentts.cpp (Qwen3-TTS-12Hz): a 28-layer Qwen3-shaped talker samples codebook-0 tokens, a 5-layer MTP code predictor fills codebooks 1–15, and the 12.5 Hz codec (RVQ dequant → sliding-window-72 transformer → ConvNeXt upsample → DAC v2) renders 1920 samples per frame at 24 kHz. CustomVoice preset speakers; CPU only.

Run¶

echo "Hello from Fucina." | zig build qwen3tts -Doptimize=ReleaseFast -- \
    --model models/qwen3-tts/qwen-talker-0.6b-customvoice-F32.gguf \
    --codec models/qwen3-tts/qwen-tokenizer-12hz-F32.gguf \
    --speaker Aiden -o hello.wav

Flags: --speaker, --lang (default english), --seed N (printed for replay) / --greedy, --max-new, --chunk-frames/--left-ctx (streamed codec decode geometry), --threads N, -o out.wav.

Models: Serveurperso/Qwen3-TTS-GGUF — one talker GGUF (qwen-talker-0.6b-customvoice-*) plus the shared codec GGUF (qwen-tokenizer-12hz-*). The F32 pair is the parity reference.

Parity¶

The port is pinned against the qwentts.cpp oracle (refs/qwentts.cpp, CPU, --no-fa): greedy generation matches token-for-token (37×16 frames, natural EOS) and seeded sampling replays the oracle draw-for-draw (Philox4x32-10, one subsequence per primitive sample); the codec decodes the golden RVQ stream at ≥0.9999 waveform cosine with per-stage probes (src/llm/qwen3tts/*_tests.zig, model-gated). On an M1 Max the 0.6B F32 talker generates ~13 fps against the 12.5 fps real-time budget and the codec decodes at ~0.5× RTF one-shot.