Capability matrix
- Compute and format FP32 · Historical Python reference on Ryzen 9 9950X3D: five-minute 7.25% WER, RTF 0.771.
- Local entry point Audio transcription.
- Server Unverified for this artifact. Check for a service that accepts this artifact.
- Harness No recorded client task. Run one identified request, then a task with a checked result.
Choose the download
Publisher model card · CKE input files
Publisher task: automatic speech recognition. CKE example: FP32. Use this artifact identifier:
hf://nvidia/parakeet-tdt-0.6b-v3.
Get started
Generated-C frontend, encoder, and TDT decoder with a native transcription host; the Python native-kernel session remains a reference path. See the Parakeet HTML guide.
version/v8/scripts/cks-v8-run audio \
hf://nvidia/parakeet-tdt-0.6b-v3 \
--wav recording.wav \
--output build/parakeet-transcript.json \
--resumeWhat has actually run?
Generated-C frontend, encoder and TDT decode under a native host; historical Python reference: exact 50-token/50-duration trajectory and complete 42-minute run. English audio; multilingual speech, concurrent sessions, and diarization remain open.
CPU evidence. Historical Python reference on Ryzen 9 9950X3D: five-minute 7.25% WER, RTF 0.771; 42-minute RTF 0.847, 3.29 GiB peak RSS. Not generated-runtime benchmarks..
Performance. For speech, compare transcription time with audio duration and check the transcript against known words. ARM throughput for this exact artifact is unverified here. Benchmark evidence · 32K to 128K results
Can I fine-tune it?
Full-checkpoint training, LoRA and other PEFT methods: no verified CKE recipe for this exact checkpoint in the reviewed evidence. For a trainable starter, use the FP32 and adapter recipes.