Lont联合创始人Ruud在Hacker News上发布了名为Overtone的开源工具，旨在通过结合多个语音转文本（STT）模型和大语言模型（LLM），实现视频字幕的自动修正与格式化 1。该工具整合了gpt-transcribe、Azure STT和Elevenlabs Scribe三种STT引擎进行转录，并利用LLM根据上下文解决冲突及纠正错误；相比单一模型使用Overtone可降低词错误率（WER） 1。
除了核心的转录校正功能，Overtone还内置了基于规则的字幕切割器，能够适配16:9、9:16等多种屏幕比例 1。该工具支持通过Video.js插件或MP4硬字幕烧录方式生成高质量视频字幕 1。项目以MIT许可证开源，提供API集成及命令行工具，相关代码仓库位于https://github.com/lonthq/overtone-captions，演示地址为https://overtone.lont.ai/ 1。
Ruud, co-founder of Lont, has released Overtone, an open-source tool designed to automatically transcribe, correct, and format video captions by combining multiple speech-to-text (STT) models with large language model (LLM) capabilities 1. The system utilizes gpt-transcribe, Azure STT, and Elevenlabs Scribe for initial transcription, employing an LLM to resolve conflicts between the outputs and apply context-aware corrections 1. By leveraging this ensemble approach, Overtone achieves a lower Word Error Rate (WER) compared to using any single model in isolation 1.
Beyond accuracy improvements, the tool features a rule-based caption cutter that adapts text layout to various screen aspect ratios, including 16:9 and 9:16 1. Users can integrate the output via Video.js plugins or burn hard subtitles directly into MP4 files 1. Overtone is available under the MIT license and provides both API integration and command-line tools for generating high-quality captions 1.
评论
还没有评论，欢迎留下第一条。