qwen3-tts-voices-GGUF

Qwen3-TTS Base voice packs for CrispASR

The new *-designed.gguf files use one freshly AI-generated VoiceDesign reference, encoded separately with the matching original Base model. No human recording, named-person imitation request, preset speaker ID or third-party voice prompt was supplied. The voice is AI-generated; speaker identity is unverified, not a commissioned actor with a performer release.

The new provenance and acceptance claims apply only to the designed files. The legacy file is retained for compatibility and is outside this new audit. A 0.6B speaker embedding cannot be padded or reused as a 1.7B embedding. The new packs contain one voice named default, selected from the original clear candidate; the raw masculine candidate failed full-reference transcription and was rejected. The encoder arrays are unchanged from the original model-specific F32 extraction.

Use

Place the matching designed pack beside the talker model and the 12 Hz codec. CrispASR's updated CLI discovers the pack by the model's metadata, including for renamed models. Or choose it explicitly:

crispasr -m qwen3-tts-12hz-1.7b-base-f16.gguf \
  --voice qwen3-tts-voice-1.7b-designed.gguf --tts "Hello world."

The registry integration downloads the matching pack and license/provenance companions with -m auto --auto-download --backend qwen3-tts (0.6B) or --backend qwen3-tts-1.7b-base (1.7B). Older CrispASR versions need an explicit path or a model-matched pack installed manually; do not substitute the legacy 1024-d pack for 1.7B.

License scope and attribution

Voice assets and our licensable data contributions are offered under Apache-2.0, to the extent rights are held; see LICENSE and NOTICE. The copied reproduction recipes are separately MIT-licensed CrispASR code, with the original copyright and permission notices retained in LICENSE-RECIPE.txt. Both license texts and attribution are also embedded in each designed GGUF. Preserve required notices; keeping the full provenance bundle is recommended, not an extra license condition. No exclusive copyright in machine-generated material, endorsement by Qwen/Alibaba, identity clearance, performer release or universal non-infringement warranty is claimed. Human listening feedback remains pending. Listening cannot establish upstream training-data rights or prove absence of likeness to every real person. Any specific reported concern should be reviewed and the affected voice replaced or withheld while unresolved. Speaker identity remains unknown in the assets.

Qwen Team / Alibaba Cloud publish the following model revisions as Apache-2.0:

Full pinned publisher declarations and cards are retained in licensing/. The exact qwen-tts 0.1.1 wheel and executed package sources were hash-verified; its full Apache license and metadata are retained. These are publisher license declarations, not an independent clearance of the upstream training corpus.

Reproduction and acceptance

release-provenance.json identifies all source, build and validator commits. The package includes the saved, disclosed reference WAV, original speaker/code arrays, generation description/transcript/seed/settings, dependency versions, source file hashes and exact bake/curation recipes. Reproduce in the full CrispASR checkout at the recorded commits.

CPU run39942586420 and actual Tesla T4 CUDA run issue555-designed-voices-cuda-v1 each pass all13 reference/speech checks at WER0, plus clean rejection of the incompatible legacy pack. Tests cover 0.6B Q8, 1.7B F16/Q8, renamed-model discovery, explicit packs and C ABI synthesis. The CPU/CUDA default GGUF bytes are identical. Per-child GPU residency and native CUDA backend logs are retained; no Vulkan-hardware or performance claim is made. The individual receipts are included. These technical checks do not establish a legal warranty about third-party rights.