Skip to main content

Multimodal & Speech

Vision-language generation​

trtmc run vision-language.bundle \
  --image input.jpg \
  --prompt "Describe this image in one sentence." \
  --max-new-tokens 48

Image shape, dynamic resolution, preprocessing metadata, decoder layout, and LoRA support are family-owned. Use registered Qwen-VL configuration fields only with a Qwen-VL build that declares them.

Speech recognition​

trtmc transcribe speech-to-text.bundle \
  --audio input.wav \
  --beam-size 1 \
  --source-language en \
  --target-language en \
  --task transcribe

Add --stream and the model-supported chunk/context controls for streaming ASR. Beam size, language pairs, punctuation, timestamps, and maximum input duration remain model-owned constraints.

Audio generation and speech-to-speech​

trtmc generate-audio text-to-audio.bundle \
  --prompt "A clear short test sentence." \
  --output output.wav

trtmc speak speech-to-speech.bundle \
  --audio-in input.wav \
  --audio-out response.wav

Use exact checkpoints from Model Recipes, organized by the corresponding Hugging Face multimodal or audio task. For concept-first labs covering preprocessing, engine components, streaming state, and typed results, follow Multimodal and Speech and Canary Decoding.