Handy Computer Shrinks Parakeet Ultra to 486 MB for Python-Free Transcription
Moondream's Parakeet Ultra speech-to-text model now ships as GGUF quantizations for transcribe.cpp, delivering 1.8% WER at 486 MB across 25 languages.
- Parakeet Ultra now available as GGUF quantizations for transcribe.cpp, from F32 (2.51 GB) to Q4_K_M (486 MB).
- Q4_K_M quant hits 1.82% WER on LibriSpeech test-clean vs 1.80% for F32, near-lossless compression.
- Covers 25 European languages with FastConformer encoder and TDT transducer decoder, 0.6B parameters.
- Built-in voice-activity head segments long audio at pauses into chunks of max 30 seconds, no external VAD needed.
- Upstream Ultra cuts FLEURS 25-language WER from 11.62% to 9.55% vs base parakeet-tdt-0.6b-v3.
- CC-BY-4.0 license, 16 kHz mono WAV input, not streaming, does not translate.
Handy Computer has released six GGUF builds of Parakeet Ultra for transcribe.cpp. The port runs local speech recognition from one C++ executable and one model file, without Python, a CUDA runtime, or a cloud request. Files range from a 2.51 GB F32 build to a 486 MB Q4_K_M quantized version.
Inside the port
The upstream Parakeet Ultra model is Moondream’s post-trained version of NVIDIA’s 600 million-parameter parakeet-tdt-0.6b-v3. Ultra retains the base model’s architecture, tokenizer, parameter count, and language set while replacing its weights with versions trained to reduce word error rate.
A FastConformer encoder converts audio frames into compact representations, while a Token-and-Duration Transducer decoder predicts both text tokens and their durations. Duration prediction lets the decoder advance through stretches that produce no token. A voice-activity detection head splits long recordings at pauses into segments of up to 30 seconds. The model accepts 16 kHz mono WAV audio and produces cased, punctuated transcripts with optional token-level timestamps.
Six files, a 0.02-point spread
Word error rate, or WER, measures substitutions, deletions, and insertions against a reference transcript; lower scores are better. On all 2,620 utterances in LibriSpeech test-clean, the reported results vary by only 0.02 percentage points from F32 to Q4_K_M.
| Precision | File size | WER | Difference from F32 |
|---|---|---|---|
| F32 | 2.51 GB | 1.80% |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
