HackerNews中文版

我开始思考为什么 Fireworks 不支持语音模型。现在有很多优秀的开源模型可用，比如 parakeet、kokoro、Qwen ASR 等，但如果没有自己管理大量的 GPU，就无法使用它们。即使是像 Gemma 4 这样被语音助手使用的 LLM 也不支持。然后我意识到，推理平台需要针对你使用的用例进行不同的优化。以 LLM 为例，甚至不是 STT 和 TTS。 * 编码助手 -> 大量缓存输入，需要针对 KV 缓存进行优化 * 创建幻灯片/博客 -> 大量输出，需要针对推测性解码进行优化 * 语音 LLM -> 缓存输入，小量输出，尚未弄清楚如何优化这一点。所以 TTS 和 STT 是完全不同的领域。我不知道的是时机，人们现在就想使用像 kokoro、parakeet、Qwen 等开源模型吗？

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?

为什么 Fireworks 不支持语音 AI