Speech-to-text (STT)
AI technology that converts spoken audio into written text — also called Automatic Speech Recognition (ASR). The input half of voice AI, distinct from TTS which produces speech.
STT (or ASR) is the gateway between human speech and LLM processing. Whisper, Deepgram, AssemblyAI, and Google Speech-to-Text are the 2026 leaders. Word accuracy on clean English audio routinely exceeds 95% with proper acoustic conditions.
For voice agents specifically, STT must be streaming and low-latency: words become text within ~100ms of being spoken, so the LLM can start generating a response before the user finishes the sentence. Whisper streaming and Deepgram Nova are purpose-built for this.
The hardest STT challenges in 2026 are accents, code-switching (mid-sentence language changes), background noise, and domain-specific vocabulary. Custom models trained on your domain audio can lift accuracy meaningfully on niche use cases.
Frequently asked
What is the best STT model in 2026?+
OpenAI Whisper Large v3 for open-source / self-hosted. Deepgram Nova-3 for production streaming. AssemblyAI for the best out-of-box conversation intelligence (speakers, summarization, topics).
How accurate is STT in 2026?+
95%+ word accuracy on clean English. Drops to 85–92% with accents, noise, or domain terminology. For legal-grade transcription, AI is the first pass and humans verify.