Instructions to use Fritter-ASR/Fritter-Nemotron-0.6B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- NeMo
How to use Fritter-ASR/Fritter-Nemotron-0.6B with NeMo:
import nemo.collections.asr as nemo_asr asr_model = nemo_asr.models.ASRModel.from_pretrained("Fritter-ASR/Fritter-Nemotron-0.6B") transcriptions = asr_model.transcribe(["file.wav"]) - Notebooks
- Google Colab
- Kaggle
Fritter Nemotron 0.6B
Formerly
danielrosehill/Fritter-Nemotron-0.6B, where Fritter was a proof of concept. It moved to the Fritter-ASR organisation in November 2026, and new releases are published here.
Streaming version of Fritter, an English dictation fine-tune for technical vocabulary, including AI terms. Current version: Fritter Nemotron v1.0 (run r31-nemotron-general, 2026-11-05). It is a full fine-tune of nvidia/nemotron-speech-streaming-en-0.6b, trained on one speaker, so other voices may gain less.
Files
| Build | Where | Size |
|---|---|---|
| NeMo weights | Fritter-Nemotron-0.6B.nemo |
2.47 GB |
| GGUF F16, for transcribe.cpp | gguf/Fritter-Nemotron-0.6B-F16.gguf |
1.24 GB |
| sherpa-onnx int8, 560 ms chunks | Fritter-ASR/Fritter-Nemotron-0.6B-sherpa-onnx-int8 |
662 MB |
Direct download (sherpa-onnx int8)
The sherpa-onnx int8 build is also served outside Hugging Face, always at the latest version:
- Manifest (version, sizes, sha256): https://models.carrotcakeai.com/asr/fritter-streaming/latest/manifest.json
- https://models.carrotcakeai.com/asr/fritter-streaming/latest/encoder.int8.onnx
- https://models.carrotcakeai.com/asr/fritter-streaming/latest/decoder.int8.onnx
- https://models.carrotcakeai.com/asr/fritter-streaming/latest/joiner.int8.onnx
- https://models.carrotcakeai.com/asr/fritter-streaming/latest/tokens.txt
- https://models.carrotcakeai.com/asr/fritter-streaming/latest/bpe.vocab
Results
Not benchmarked yet. Validation WER on this run's own validation split was 5.83%. That split picks the checkpoint, so it is not a test result and can't be compared with other models.
Training data
- 858.8 minutes of one speaker's read-aloud recordings, plus 95.6 minutes of LibriSpeech-PC replay: 11,745 clips, 1,049.8 minutes with priority repeats. Source:
danielrosehill/Read-Aloud-Voice-Sentences. - 20 epochs, lr 3e-5, batch 16, one L40S, 92.7 minutes of training.
Licence. CC-BY-NC-4.0: free for non-commercial use with attribution. The base model, nvidia/nemotron-speech-streaming-en-0.6b, is by NVIDIA. Licensed by NVIDIA Corporation under the NVIDIA Open Model License: see NOTICE and NVIDIA-Open-Model-License.pdf. A personal project by Daniel Rosehill; not affiliated with NVIDIA.
- Downloads last month
- 88
16-bit
Model tree for Fritter-ASR/Fritter-Nemotron-0.6B
Base model
nvidia/nemotron-speech-streaming-en-0.6b