Mistral 发布 Voxtral TTS：开源语音生成新突破Mistral Releases Voxtral TTS: Open-Source Speech Generation Breakthrough
关键要点：企业依赖闭源模型而非在自家专有数据上微调，正在错失巨大价值——Mistral 的新开源模型和 Forge 平台让这件事变得简单且性能大幅提升。
Mistral 首席科学家 Guillaume Lample 和音频研究负责人 Pavan Kumar Reddy 宣布了 Voxtral TTS，这是他们首个文本转语音模型。该 3B 参数模型支持九种语言，采用新型自回归流匹配架构，并搭配内部神经音频编解码器，在竞争对手成本一小部分的情况下实现最先进质量。
他们强调，企业积累了数万亿特定领域 token，但闭源模型永远无法访问这些数据。“如果使用闭源模型，他们基本上无法从这些多年来收集的所有洞见和数据中获益。”Mistral 的 Forge 平台让客户使用内部工具在自家数据上微调，实现更优、私有且定制的结果。他们还介绍了 Leanstral 用于可验证推理，以及他们推动开源以民主化智能的使命。
Mistral 首席科学家 Guillaume Lample 和音频研究负责人 Pavan Kumar Reddy 宣布了 Voxtral TTS，这是他们首个文本转语音模型。该 3B 参数模型支持九种语言，采用新型自回归流匹配架构，并搭配内部神经音频编解码器，在竞争对手成本一小部分的情况下实现最先进质量。
他们强调，企业积累了数万亿特定领域 token，但闭源模型永远无法访问这些数据。“如果使用闭源模型，他们基本上无法从这些多年来收集的所有洞见和数据中获益。”Mistral 的 Forge 平台让客户使用内部工具在自家数据上微调，实现更优、私有且定制的结果。他们还介绍了 Leanstral 用于可验证推理，以及他们推动开源以民主化智能的使命。
The Takeaway: Enterprises are leaving massive value on the table by relying on closed models instead of fine-tuning on their own proprietary data—Mistral's new open models and Forge platform make that easy and dramatically more performant.
Mistral Chief Scientist Guillaume Lample and Audio Research lead Pavan Kumar Reddy announced Voxtral TTS, their first text-to-speech model. The 3B parameter model supports nine languages and uses a novel auto-regressive flow matching architecture paired with an in-house neural audio codec delivering state-of-the-art quality at a fraction of competitors' cost.
They stressed that companies sit on trillions of domain-specific tokens that closed models never see. "If they're using like closed source models they are basically not benefiting from all this insights, all these data they have collected through years." Mistral's Forge lets customers fine-tune on their data using internal tools for superior, private results. They also highlighted Leanstral for verifiable reasoning and their open-source mission to democratize intelligence.
查看原文 →
Mistral Chief Scientist Guillaume Lample and Audio Research lead Pavan Kumar Reddy announced Voxtral TTS, their first text-to-speech model. The 3B parameter model supports nine languages and uses a novel auto-regressive flow matching architecture paired with an in-house neural audio codec delivering state-of-the-art quality at a fraction of competitors' cost.
They stressed that companies sit on trillions of domain-specific tokens that closed models never see. "If they're using like closed source models they are basically not benefiting from all this insights, all these data they have collected through years." Mistral's Forge lets customers fine-tune on their data using internal tools for superior, private results. They also highlighted Leanstral for verifiable reasoning and their open-source mission to democratize intelligence.