Fish Audio Pursues Live Conversation Stack

Fish Audio is trying to move from being a voice output vendor to owning the live conversation stack. Today its distribution is still anchored in text-to-speech integrations, like LiveKit plugins and Telnyx voice catalog support, but speech-to-speech and audio understanding would let it process the raw call itself, detect pauses and interruptions, preserve tone, and sell a larger share of each production voice workflow.1 2 3 4 5

• The practical shift is from generating the final spoken reply to handling more of the loop. LiveKit already handles turn detection, interruptions, and orchestration around Fish Audio, which shows where value sits today and what Fish Audio would need to internalize to become a control layer instead of a plug in voice model.2 5
• Comparable voice infrastructure companies that own broader speech workflows are already larger. Deepgram sells speech-to-text, text-to-speech, and real-time voice understanding, and is estimated at $100M revenue as of July 2026. Retell, which sits at the agent orchestration layer, reached $60M annualized revenue in April 2026. Fish Audio was at $21M annualized revenue in July 2026.3 4 6
• The competitive pressure is that native speech-to-speech models can collapse the old ASR, LLM, and TTS chain into one realtime model call. OpenAI positions Realtime this way, with built in speech-to-speech and server side turn detection, which means Fish Audio needs deeper workflow ownership, not just better voice quality, to protect pricing power.1 7 8

The next step is a market split between interchangeable voice components and sticky voice control planes. If Fish Audio keeps expanding from expressive synthesis into turn taking, call monitoring, and audio understanding, it can move closer to the Retell and Deepgram part of the stack, where the vendor sees more of the call, captures more data, and becomes harder to replace.3 4 5 7