为什么Fireworks不支持语音人工智能
我开始思考为什么火花不支持语音模型。目前有很多优秀的开源模型可用，比如parakeet、kokoro、Qwen ASR等，但如果不自己管理一堆GPU，就无法使用它们。
即使是像Gemma 4这样的LLM（大型语言模型），在语音代理中也不被支持。
然后我意识到，推理平台需要根据你所使用的用例进行不同的优化。以LLM为例，而不仅仅是语音转文本（STT）和文本转语音（TTS）：
- 编码代理 -> 需要大量缓存输入，需要针对KV缓存进行优化
- 创建幻灯片/博客 -> 输出量大，需要针对推测解码进行优化
- 语音LLM -> 缓存输入小，输出少，目前还没有找到优化的方法。
因此，TTS和STT是完全不同的领域。
我不知道的是时间问题，大家现在是否真的想使用像kokoro、parakeet、Qwen等开源模型？
查看原文
I started thinking over why doesn't fireworks support voice models. There are really good opensource models available now, like parakeet, kokoro, Qwen ASR etc but no way to use it without managing a bunch of GPUs yourself.<p>Even LLMs like Gemma 4 used by voice agents are not supported.<p>Then I figured that the inference platform needs to be optimized differently for the kind of usecase you are using. Lets take an example for LLMs, not even STT and TTS.
- Coding agents -> lot of cached input, needs to optimize for KV cache
- Creation slides/blogs -> lots of output, needs to optimize for speculative decoding
- Voice LLMs -> Cached input small output, not yet figured out on how to optimize this.<p>So TTS and STT is a completely different ballgame.<p>What I don't know is the timing, do people want to use open source models like kokoro, parakeet, Qwen etc RIGHT NOW?