Fritter models
Speech models for English dictation, fine-tuned on one Irish-English speaker's voice: AI model names, Linux and developer vocabulary, brand names, acronyms, and numbers written as digits. Everything you need to ship one: which model fits, direct links to the weights, the licence, and the line to credit it with.
Collections: Fritter-ASR-Parakeet · Fritter-ASR-Nemotron · Fritter TTS
Speech recognition, offline (whole clips)
Fine-tune of NVIDIA Parakeet TDT 0.6B v2 for English dictation: AI model and product names, Linux and developer vocabulary, brands, acronyms, numbers as digits. Best accuracy when you transcribe a finished recording.
Android, iOS and desktop apps with sherpa-onnx (offline transducer, model type nemo_transducer). Repo: Fritter-ASR/Fritter
Browser (parakeet.js with ONNX Runtime Web, as the demo does) and onnx-asr. Repo: Fritter-ASR/Fritter
Same runtimes as int8, full precision. Repo: Fritter-ASR/Fritter
transcribe.cpp and other ggml apps. Q4_K_M is the smallest, F16 the most exact. Repo: Fritter-ASR/Fritter
PyTorch with NVIDIA NeMo: GPU servers and further fine-tuning. Repo: Fritter-ASR/Fritter
The sherpa-onnx int8 files at a stable address for apps that update themselves. The manifest lists each file with size and sha256, and says whether the model is offline or streaming.
import nemo.collections.asr as nemo_asr
from huggingface_hub import hf_hub_download
path = hf_hub_download("Fritter-ASR/Fritter", "Fritter.nemo")
model = nemo_asr.models.ASRModel.restore_from(path)
print(model.transcribe(["clip.wav"])[0].text) # 16 kHz monoPut this (or the same facts) where your app lists its credits or licences:
Fritter by Daniel Rosehill (https://huggingface.co/Fritter-ASR/Fritter), licensed CC BY-NC 4.0, fine-tuned from NVIDIA Parakeet TDT 0.6B v2 (CC BY 4.0).Speech recognition, streaming (live captions, push-to-talk)
The streaming version of Fritter, fine-tuned from NVIDIA Nemotron streaming 0.6B on the same recordings. Text appears while you speak.
Android, iOS and desktop apps with sherpa-onnx (online transducer), 560 ms chunks. Repo: Fritter-ASR/Fritter-Nemotron-0.6B-sherpa-onnx-int8
transcribe.cpp and other ggml apps. Repo: Fritter-ASR/Fritter-Nemotron-0.6B
PyTorch with NVIDIA NeMo: GPU servers and further fine-tuning. Repo: Fritter-ASR/Fritter-Nemotron-0.6B
The sherpa-onnx int8 files at a stable address for apps that update themselves. The manifest lists each file with size and sha256, and says whether the model is offline or streaming.
Put this (or the same facts) where your app lists its credits or licences:
Fritter Nemotron by Daniel Rosehill (https://huggingface.co/Fritter-ASR/Fritter-Nemotron-0.6B), licensed CC BY-NC 4.0, fine-tuned from NVIDIA Nemotron streaming 0.6B. Licensed by NVIDIA Corporation under the NVIDIA Open Model License.Ship the NOTICE file and the NVIDIA Open Model License with any copy of these weights.
Text-to-speech
A text-to-speech voice trained on the same recordings: one Irish-English speaker reading technical and general English. No Hebrew.
Transformers / the Qwen3-TTS reference code. Download the whole repo (it includes the speech tokenizer). Repo: Fritter-ASR/Fritter-TTS
qwentts.cpp. Needs both the talker and the tokenizer file at the same quantisation. Repo: Fritter-ASR/Fritter-TTS-GGUF
Put this (or the same facts) where your app lists its credits or licences:
Fritter TTS by Daniel Rosehill (https://huggingface.co/Fritter-ASR/Fritter-TTS), licensed CC BY-NC 4.0, fine-tuned from Qwen3-TTS-12Hz-1.7B-Base (Apache 2.0).All Fritter weights are released under Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0): free for non-commercial use, with attribution.
NOTICE file and the licence with the weights), Qwen3-TTS (Apache 2.0).Fritter is a personal project by Daniel Rosehill, not affiliated with NVIDIA or Qwen. This page is a summary, not legal advice; the licence texts govern. Questions and release notes: github.com/Fritter-ASR/Fritter.