做语音分析或实时内容审核的开发者，Velma 直接跳过转录环节，成本低且能捕捉语气情绪，值得一试。
开发者 Santiago 在构建语音管道时发现，传统方法将音频转为文本后丢失了语气、犹豫、讽刺等关键信息。Modulate AI 的 Velma 模型直接处理原始音频，无需转录，能检测多达 150 种隐形线索，已在《使命召唤》和《GTA Online》中用于实时检测毒性言论。Velma 通过 API 提供，成本比通过 LLM 处理音频低约 10 倍。该模型解决了语音分析中语义与情感分离的痛点，适合需要深度理解语音意图的团队。
I've built two voice pipelines for two different c…
I've built two voice pipelines for two different companies.
They both look like this:
Audio → STT → Clean transcript → NLP → Classify → Act
This works, but there's still a problem I can't solve.
Every time I convert audio to text, I'm keeping the words but throwing away the meaning. Tone, hesitation, sarcasm, and stress are all gone. I have the text, but miss its soul.
The folks at @modulate_ai reached out and showed me how to solve this.
Velma is the voice model that's been running inside Call of Duty and GTA Online to catch toxicity in real time.
This model skips the transcript entirely and works directly on the raw audio. This allows the model to take into account the "invisible clues" other models miss.
It can detect up to 150 invisible clues that none else does!
You can access Velma through an API, and it's ~10x cheaper than pushing audio through an LLM.
If you want to give it a try, use this link to get 1,000 free credits:
https://t.co/oVafRuuKBg
Thanks to the team for partnering with me on this post.