About This Project

Fritter is a specialized fine-tune of the NVIDIA Parakeet TDT 0.6B v2 model, engineered to improve speech-to-text accuracy for a specific speaker. By training on a custom dataset of Irish-English phone recordings, the model achieves significant performance gains in recognizing technical terminology, cloud computing jargon, and brand names, while maintaining natural phrasing in everyday email and message dictation.

The project leverages the NeMo framework for training, resulting in a model that significantly reduces Word Error Rate (WER) for its target speaker compared to the base model. It is designed for users requiring high-accuracy dictation for technical workflows and supports deployment via ONNX and GGUF formats for efficient, local, offline use on devices. It is particularly well-suited for developers and power users who need a reliable, personalized transcription tool that handles specific professional vocabulary without the latency of cloud-based services.

Project Details

Tags

nemoonnxggufparakeettdtspeaker-adaptationenglishtech-vocabularyirish-englishdictationfritterautomatic-speech-recognitionendataset:danielrosehill/read-aloud-voice-sentencesdataset:openslr/librispeech_asrbase_model:nvidia/parakeet-tdt-0.6b-v2base_model:finetune:nvidia/parakeet-tdt-0.6b-v2doi:10.57967/hf/11468license:cc-by-nc-4.0region:us

Explore More Projects