English dictation models for technical vocabulary, including AI terms: model and product names, CLI and cloud terms, brands, acronyms and numbers in digits. By Daniel Rosehill.
Stock speech models stumble on the words technical people say all day. I fine-tuned Fritter on my own read-aloud recordings of that vocabulary, so it is trained on one speaker: me. Other voices may gain less.
On held-out recordings of my voice, Fritter v1.0 took tech-sentence word error rate from 7.53% to 4.01%. Later versions have not been benchmarked, and general English came out slightly worse than stock.
Licence: CC-BY-NC-4.0, free for non-commercial use with attribution. The streaming build's base model is by NVIDIA: licensed by NVIDIA Corporation under the NVIDIA Open Model License (notice, agreement). A personal project, not affiliated with NVIDIA.
Each link serves the latest version in sherpa-onnx int8. The manifest gives the version, and each file's size and sha256.
v1.2 · 661 MB · by Daniel Rosehill
Offline: transcribes a finished recording.
Base model: nvidia/parakeet-tdt-0.6b-v2
Trained on: Fine-tuned from stock on 694.8 minutes of my own recordings plus 77.3 minutes of LibriSpeech replay (v1.1), then continued on 210.0 minutes of new recordings plus a replay sample (v1.2).
On Hugging Face:
Fritter Nemotron v1.0 · 662 MB · by Daniel Rosehill
Streaming: writes text as you speak, in 560 ms chunks.
Base model: nvidia/nemotron-speech-streaming-en-0.6b
Trained on: Fine-tuned from stock on 858.8 minutes of my own recordings plus 95.6 minutes of LibriSpeech replay. Not benchmarked yet.
On Hugging Face:
