Destined AI公司在Hacker News上开源了Hurston-1文本转语音模型，采用Apache-2.0许可证1。该模型专门为美国南方口音和非洲美国英语优化，是一个窄带TTS模型，主要面向电话应用等领域1。
Hurston-1基于大规模语音数据微调而成1。该公司自2021年以来收集了来自20,000多人的全知情同意语音数据，累计超过1,800小时、100万条音频录音，其中136小时用于模型微调1。这家总部位于亚特兰大的公司已将该模型应用于Amazon和Salesforce等客户，商业化运营已产生超过50万美元收入1。
模型具有一定的技术限制。开发者说明，Hurston-1不能保持跨提示的完全相同声音，也不支持声音克隆功能1。
Destined AI has open-sourced Hurston-1, a text-to-speech model optimized for Southern American accents and African American English, under the Apache-2.0 license 1. The narrowband TTS model, designed primarily for telephone applications, was fine-tuned on 136 hours of audio data collected from more than 20,000 speakers who provided full informed consent since 2021 1. The underlying training dataset comprises over 1.8 million audio recordings totaling more than 1,800 hours 1.
The model has already demonstrated commercial viability, having supported customers including Amazon and Salesforce and generating over $500,000 in revenue 1. However, the system has specific limitations: it cannot maintain consistent voice characteristics across different prompts and does not support voice cloning 1. Destined AI, based in Atlanta, announced the open-source release on Hacker News 1.
评论
还没有评论，欢迎留下第一条。