Overtone releases tool to cut captions for any screen shape

The system combines multiple speech engines, including OpenAI and Azure AI Speech, to improve accuracy. It uses an LLM to correct misheard words and names based on context, while ensuring timestamps remain accurate by relying on the speech engines for timing. Benchmarks show Overtone achieves a 5.3% word error rate, outperforming individual engines like gpt-transcribe (6.2%) and Azure Speech (7.7%). It also significantly reduces line breaks ending on weak words, with only 0.3% of breaks ending on words like "the" or "of", compared to 28.8% for standard line-filling methods.

The tool is designed for offline use without API keys for recutting existing subtitles, though full transcription requires provider keys. It supports languages including English, Dutch, German, French, Spanish, Italian, and Portuguese, with specific rules for line breaking and number handling. The project is public, with full method details and caveats available in its documentation.