Qwen3-TTS-12Hz-0.6B-CustomVoice — OpenVINO IR
A ready-to-use OpenVINO IR export of Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice, prepared for Fern. The export preserves the upstream model weights and architecture; it is not fine-tuned.
What this variant does
Generate speech with the model’s built-in speaker choices and supported style instructions. This variant does not include the Base model’s reference-audio cloning mode.
- Export size: approximately 2.025 GiB of model files.
- Representation: OpenVINO IR with FP16-compressed constants. No NNCF quantization is applied.
- Files: model XML/BIN pairs, processor and tokenizer assets, and both OpenVINO speech-tokenizer components.
fern-export-manifest.json records each exported file’s size and SHA-256 digest, plus the source-cache revision information.
- Not included: original PyTorch checkpoints, local caches, voice recordings, or generated audio.
Use with Fern
Install the matching Qwen3-TTS model from Fern’s audio workspace. Fern downloads this ready-made export; users do not need to convert the source checkpoint. The first load compiles the IR for the selected device and can take longer than later loads.
Python example
The following uses Fern’s Qwen3-TTS helper. Run it in an environment with the Fern scripts/tts_runtime directory on PYTHONPATH and the runtime dependencies listed below.
from huggingface_hub import snapshot_download
from qwen_3_tts_helper import OVQwen3TTSModel
model_dir = snapshot_download(
repo_id="xgirls/Qwen3-TTS-12Hz-0.6B-CustomVoice-OpenVINO",
revision="56999806c869a759c3923de3f2ffc6021c94d582",
)
model = OVQwen3TTSModel.from_pretrained(model_dir=model_dir, device="GPU")
Runtime and validation
Fern’s integration uses the Qwen3-TTS helper at source commit 1ab0dd75353392f28a0d05d9ca960c9954b13c83. The validated environment used OpenVINO 2026.4.0, Transformers 4.57.3, PyTorch and torchaudio 2.8.0, and Qwen-TTS 0.0.4. This is an application-specific runtime set, not a claim that every combination of library versions or OpenVINO devices is supported.
Fern API smoke validation generated non-silent 24 kHz audio with this model variant. This verifies basic generation, not comparative quality or performance. First-load compilation is device-specific.
License and attribution
Apache-2.0; see the included LICENSE and NOTICE. The original model is from the Alibaba Qwen team. This community OpenVINO conversion is not an official Qwen or Intel release. The export follows the OpenVINO Qwen3-TTS notebook workflow.
