Skip to content

Parakeet

Parakeet TDT V3 via onnx-asr.

Run hyprwhspr setup and select Parakeet → Parakeet v3. The model (~1 GB) is downloaded during setup.

{
"transcription_backend": "onnx-asr",
"onnx_asr_model": "nemo-parakeet-tdt-0.6b-v3", // default
"onnx_asr_quantization": "int8", // int8 (default) | null (FP32)
"onnx_asr_use_vad": true, // Silero VAD for longer recordings (default: true)
"onnx_asr_vad_min_duration": 30, // seconds before VAD is used (default: 30)
"onnx_asr_vad_batch_size": 1 // VAD segments per pass (default: 1)
}

Keep onnx_asr_vad_batch_size at 1 to limit memory use. Larger batches can speed up long GPU recordings, but use more memory. The setting applies when VAD is enabled and the recording reaches onnx_asr_vad_min_duration.

The service log reports the execution provider in use, including CPU fallback. For automatic unloading between recordings, see Memory management.

Orukeet: an optional Parakeet v3 fine-tune for 25 European languages. Pick Parakeet → Orukeet in setup, or:

Terminal window
hyprwhspr setup auto --backend onnx-asr --model orukeet

Or set "onnx_asr_model": "orukeet". Always int8; your onnx_asr_quantization stays put for Parakeet.

About 672 MB, checked against pinned checksums, run locally. On a checksum error, hyprwhspr model download repairs the file.

Benchmarks and limitations.