dabri-models / README.md
AshBuk's picture
Add Q4_K accuracy to the model card
8d2a924 verified
|
Raw History Blame Contribute Delete
3.15 kB
---
license: cc-by-4.0
base_model: moondream/parakeet-ultra
base_model_relation: quantized
library_name: whisper.cpp
pipeline_tag: automatic-speech-recognition
language: [en, de, fr, es, it, pt, ru, uk, hr, sl, lv, lt, et, fi, sv, da, nl, pl, cs, sk, hu, ro, bg, el, mt]
tags: [parakeet, tdt, ggml, whisper.cpp]
---
# Dabri models
Speech models converted for [Dabri](https://github.com/AshBuk/dabri), an offline speech-to-text desktop app for Linux.
## Parakeet Ultra
[Moondream Parakeet Ultra](https://huggingface.co/moondream/parakeet-ultra) in the ggml format of libparakeet from [whisper.cpp](https://github.com/ggml-org/whisper.cpp) v1.9.5, the same format as [ggml-org/parakeet-GGUF](https://huggingface.co/ggml-org/parakeet-GGUF). Parakeet Ultra is Moondream's post-trained [nvidia/parakeet-tdt-0.6b-v3](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) with the same architecture and tokenizer. It covers the same 25 European languages and detects the language itself.
Dabri downloads `ggml-parakeet-ultra-q4_k.bin` (416 MB, SHA-256 `c463493169bd63c4de3787357f1c41f1cd51f83fb0b05a56bf64990fcc82f2f6`) when you pick the model in the tray or run `dabri model set parakeet-ultra-q4_k`. `ggml-parakeet-ultra-q8_0.bin` (669 MB, SHA-256 `fa28ec8c6afd9893da9829ec9f2fa908a60bc0fbebd37fd60bbd2f0953076fb2`) is the larger option. Both run with whisper.cpp's `parakeet-cli`:
```bash
parakeet-cli -m ggml-parakeet-ultra-q4_k.bin -f audio.wav
```
### Accuracy
Measured on LibriSpeech test-clean (2620 utterances) with whisper.cpp's `tests/librispeech-parakeet` normalizer and `parakeet-cli` on the CPU. The F16 conversion scores 1.80% WER, the figure transcribe.cpp reports for Moondream's own runtime on this split, and Q4_K scores 1.81%. In the same setup `ggml-parakeet-tdt-0.6b-v3-q8_0.bin` scores 1.95%. Q8_0 matched F16 on the 776 utterances it was run on (1.45% each).
### How the files were made
`convert-parakeet-ultra.py` in this repository needs only NumPy. Moondream ships HF safetensors, so the script takes `ggml-parakeet-tdt-0.6b-v3-f16.bin` from ggml-org/parakeet-GGUF as a template: the header, mel filterbank, window, TDT durations and vocabulary stay as they are, and each tensor is replaced by the Ultra tensor of the same name. The HF to NeMo name mapping comes from [transcribe.cpp](https://github.com/handy-computer/transcribe.cpp) (MIT). Ultra's voice-activity head has no counterpart in libparakeet and is left out, so audio longer than 30 seconds is not split by voice activity the way Moondream's runtime does it.
```bash
hf download moondream/parakeet-ultra --revision 73175eb7aeb0d82f1e2a6b53b3aabc10a90bcd0b --local-dir parakeet-ultra
hf download ggml-org/parakeet-GGUF ggml-parakeet-tdt-0.6b-v3-f16.bin --local-dir .
python convert-parakeet-ultra.py ggml-parakeet-tdt-0.6b-v3-f16.bin parakeet-ultra ggml-parakeet-ultra-f32.bin
parakeet-quantize ggml-parakeet-ultra-f32.bin ggml-parakeet-ultra-q4_k.bin q4_k
parakeet-quantize ggml-parakeet-ultra-f32.bin ggml-parakeet-ultra-q8_0.bin q8_0
```
## License
The models are licensed CC-BY-4.0 by Moondream, the same as the original. The conversion script is MIT.