Inherited NVIDIA research scripts

These scripts are retained for engine investigation, separate from product performance qualification. Run them inside the Docker runtime with NVIDIA GPU access and this directory mounted at /benchmarks. The installed model and faster packages both come from this project.

Clone benchmarks require an authorized reference recording mounted in the container. Set QWEN_TTS_REF_AUDIO to its container path and QWEN_TTS_REF_TEXT to its exact transcript. There is no default recording, transcript download or local ASR dependency. custom_voice.py uses a built-in speaker instead. Mount a writable output directory if saving generated WAVs.

For example, after building the image and preparing permitted assets in a local benchmark-data directory:

docker run --rm --gpus all \
  -v faster_qwen3_tts_studio_data:/app/persistent \
  -v ./benchmarks/research:/benchmarks:ro -v ./benchmark-data:/data -w /data \
  -e QWEN_TTS_REF_AUDIO=/data/reference.wav \
  -e QWEN_TTS_REF_TEXT="The exact words in your reference recording." \
  --entrypoint python faster-qwen3-tts-studio:dev /benchmarks/chunk_sweep.py

Generated WAVs go to the working directory, or QWEN_TTS_BENCH_OUTPUT. Preserve measurement settings, hardware, warmup policy and the exact candidate identity before comparing results. The historical tables are upstream measurements, not results for this Docker build. Use the maintained product benchmark suite for product comparisons.