pszemraj/parakeet-tdt-0.6b-v3-gguf

Quantized via the updated crispasr-quantize tool. This CLI provides superior performance over the older cohere-quantize tool due to better tensor handling and model-agnostic alignment logic. Empirical tests confirm higher quality compared to versions using the legacy tool, specifics are in the manifest.json

quantized versons of nvidia/parakeet-tdt-0.6b-v3 in gguf format via CrispASR (ggml backend)

• made for use with parakit, a daemon-style local dictation tool I made
• Empirical tests comparing the Q8, the F16, and the original Nemo file (all in this repo) on a ~60s dictation show essentially no degradation in both Q8 and F16