Freestyle OSS adds ASR biasing for smarter voice dictation
Guide words like 'Knicks' and 'OG Anunoby' now appear correctly in transcriptions.
A developer known as matt8p has open-sourced Freestyle, a voice dictation app inspired by Wispr Flow. Over the past few weeks, he built basic speech-to-text capabilities and then tackled one of Wispr Flow's standout features: ASR biasing.
ASR (automatic speech recognition) biasing is a technique that gives the transcription model hints about how words are spelled or which phrases are common. In practice, if you frequently talk about the 'Knicks' or 'OG Anunoby,' you can add those terms to your vocabulary. The model then prioritizes those words when it hears similar-sounding audio. Implementation varies by provider: OpenAI and Groq use a 'prompt' parameter (like an LLM system prompt), while Deepgram and Eleven Labs use dedicated 'key terms' parameters. Freestyle stores your vocabulary locally and injects it fresh into every inference call. The entire project is available on GitHub, and the developer is seeking contributors to advance open-source voice dictation.
- Freestyle implements ASR biasing by injecting user-defined vocabulary into the model's prompt or key terms for each transcription request.
- Supported providers include Groq (whisper-large-v3-turbo), OpenAI, Deepgram, Eleven Labs, and local models like whisper.cpp and MLX.
- The feature is fully open-source on GitHub; the project aims to build a community around voice dictation tools.
Why It Matters
Open-source ASR biasing makes voice dictation more accurate for domain-specific terms, benefiting developers and power users.