Acknowledgements
What Fraze is built on.
The Fraze desktop app ships the open-source software and models below, each under its own license. This list covers what actually ships; it does not cover this website.
Bergamot translator and Firefox Translations models
MPL-2.0The on-device translation engine, run as WebAssembly in a background thread. The translation model files are the ones Firefox ships, from Mozilla's Firefox Translations project.
OpenAI Whisper base
MITThe model the free audio language detector runs in your browser, loaded with transformers.js (Apache-2.0).
sherpa-onnx
Apache-2.0Runs speech recognition (the speech models below) and voice activity detection on-device.
NVIDIA Parakeet TDT 0.6B v3
CC-BY-4.0The speech recognition model for the 25 European languages, per its model card. Fraze downloads a quantized copy of these weights on first use; see the model card for the license that governs them.
NVIDIA Parakeet TDT-CTC 0.6B Japanese
CC-BY-4.0The speech recognition model for Japanese, per its model card. Fraze downloads a copy of these weights the first time it needs them; see the model card for the license that governs them.
SenseVoice Small
FunASR Model LicenseBy FunAudioLLM, Alibaba Group. One of the speech recognition models Fraze uses for Chinese, Korean and Cantonese. Fraze downloads a copy of these weights the first time it needs them, under the FunASR Model License, which asks that the source and the model name be kept.
Qwen3-ASR 0.6B
Apache-2.0By the Qwen team, Alibaba. One of the speech recognition models Fraze uses for Chinese, Korean and Cantonese. Fraze downloads a copy of these weights the first time it needs them.
Zipformer Vietnamese speech recognition (viet_iter3_pseudo_label)
Apache-2.0By zzasdf, trained on about 70,000 hours of Vietnamese. The speech recognition model for Vietnamese. Fraze downloads the copy converted for sherpa-onnx (csukuangfj/sherpa-onnx-zipformer-vi-int8-2025-04-20 on Hugging Face) the first time it needs it.
Zipformer Thai speech recognition (icefall-asr-gigaspeech2-th-zipformer-2024-06-20)
Apache-2.0By yfyeung, trained on the Thai part of GigaSpeech 2. The speech recognition model for Thai. Fraze downloads it from Hugging Face the first time it needs it.
Silero VAD
MITDetects when someone is speaking, so Fraze knows where to start and end a caption.
opencc-js
MITWrites Cantonese captions in Traditional Chinese characters. Its conversion dictionaries come from the OpenCC project, under Apache-2.0.