Qwen3-TTS Base 1.7B: BookAlive native ONNX conversion
First-party Nile Academy conversion of the official Qwen checkpoint at
fd4b254389122332181a7c3db7f27e918eec64e3. This is a derivative ONNX artifact,
not an ONNX release published by Qwen. It contains data only; Python is not required by BookAlive.
The fp32 pack includes a speaker encoder, talker with an explicit KV cache, a complete
code-predictor frame, speech decoder, embedding tables, tokenizer and publisher licence.
BookAlive reuses its Qwen VoiceDesign generation implementation for the Base talker,
predictor and decoder. The source implementation is pinned at
022e286b98fbec7e1e916cb940cdf532cd9f488e.
The current adapter uses x-vector-only reference cloning. It supports ten languages and
sampling temperature; transcript/codec in-context prompting and arbitrary style instructions
are not enabled. X-vector cloning follows timbre more strongly than reference rhythm. Use a
clear, authorized reference and listen to a short preview before committing to a chapter.
No automatic or guaranteed breathing control is claimed.
build_metadata.json records parent file hashes, conversion source hashes, validated graph
hashes and numerical qualification. The chained greedy check requires 48 identical codec
frames. Speaker/logit relative error must remain below 0.001 and decoder waveform absolute
error below 0.0001. These deterministic checks are separate from perceived naturalness.
Workstation development build; unsigned manifest; no human listening review or commercial
release qualification. CPU is the qualified numerical backend for this build. No RTX 4070
performance claim is made. Use reference recordings only with the speaker's permission.
Official implementation: https://github.com/QwenLM/Qwen3-TTS
