Qwen3 TTS Samples

Command-line sample

bin\x64\Release\nvigi.tts.qwen.exe demonstrates direct text synthesis, voice cloning, batch generation, and interactive streaming.

Basic synthesis

nvigi.tts.qwen.exe ^
  --models ..\..\..\data\nvigi.models ^
  --text "Qwen3 is ready for real-time speech." ^
  --output basic.wav ^
  --backend cuda

Voice cloning from a locally provisioned JSON

First run voice_cloning\provision_qwen_example_voice.ps1 from the pack root. Then:

nvigi.tts.qwen.exe ^
  --models ..\..\..\data\nvigi.models ^
  --speaker ..\..\..\data\nvigi.test\qwen_voice_refs\qwen_example_clone_2.json ^
  --text "This line uses the locally provisioned reference voice." ^
  --output clone.wav ^
  --backend d3d12

--speaker accepts a Qwen voice JSON. It is mutually exclusive with --refAudio.

Interactive streaming

nvigi.tts.qwen.exe ^
  --models ..\..\..\data\nvigi.models ^
  --speaker ..\..\..\data\nvigi.test\qwen_voice_refs\qwen_example_clone_2.json ^
  --interactive ^
  --backend vulkan

The sample streams audio chunks to the default Windows audio device. Enter exit, quit, or an empty line to leave interactive mode.

For a long paragraph in interactive or batch mode, select a Qwen --speaker JSON (or provide --refAudio and --refText). The CLI keeps that conditioning for every bounded segment. Without a voice reference, only a single short Base request is supported; a long request returns an actionable error instead of joining independently generated voices.

Batch synthesis

Use a JSONL file with one object per line. out and text are required; lang, speaker, refAudio, and refText are optional. Per-row refAudio/refText override the corresponding CLI options. Use either a Qwen speaker JSON or refAudio for a row, not both.

{"out":"english.wav","text":"Hello from Qwen.","lang":"English"}
{"out":"japanese.wav","text":"Qwen3 TTS sample.","lang":"Japanese","speaker":"..\\..\\..\\data\\nvigi.test\\qwen_voice_refs\\qwen_example_clone_2.json"}

nvigi.tts.qwen.exe --models ..\..\..\data\nvigi.models --batch jobs.jsonl --backend cuda

3D sample

bin\x64\Release\nvigi.3d.exe is an interactive D3D12 application with Qwen TTS controls. Start it after running setup_sample.bat.

nvigi.3d.exe
nvigi.3d.exe -vk

The 3D sample starts with the unconditioned Base voice and discovers Qwen voice JSONs or WAVs under data\nvigi.test\qwen_voice_refs. The provisioning script writes its generated JSON there automatically.

The standalone package validates Qwen TTS in the UI. ASR/GPT tabs require a Developer SDK overlay and are outside the standalone TTS pack scope.