From Transcription to Voice Infrastructure

Whisper turned transcription from a differentiated product into a cheap model call, which forced speech companies to move their moat somewhere else. Once a strong open source model shipped in September 2022 and the hosted API landed at $0.006 per minute, basic batch transcription became easy to swap in, easy to self host, and hard to mark up. That changed competition from who can transcribe best on average to who can win on latency, workflow fit, and full voice stack breadth.1 2 3 4

• The economic reset was severe. Deepgram had been charging about $0.0125 to $0.0145 per minute for its Base and Enhanced models, so OpenAI set a new reference price at roughly half that level. Human transcription services like Rev stayed far higher because they sell labor and review time, not just model inference.3 4
• Whisper mattered not just because it was cheap, but because it was open. A startup could wrap it into an API, an enterprise like Zoom could fine tune or self manage its own stack, and adjacent audio companies could add speech to text without building years of core research first. That is what commoditization looks like in practice.1 2 4
• The winners after Whisper look different. Deepgram pushed toward real time voice infrastructure for agents, while ElevenLabs expanded from text to speech into a broader audio platform. Their recent scale reflects that shift, with Deepgram at an estimated $100M revenue run rate in July 2026 and ElevenLabs at $600M in June 2026.5 6

The next layer of value sits above raw transcription. As voice agents spread, the defensible product is the system that hears speech fast, knows when a speaker has finished, plugs into telephony and workflows, and ships reliably at scale. Basic transcription remains abundant. Real time orchestration and packaged voice outcomes are where pricing power is heading.4 5 6