Deepgram 评测

Deepgram is an API platform for voice AI. It offers three core products: speech-to-text that converts audio to text with industry-leading accuracy, text-to-speech that generates natural-sounding voices from text, and a Voice Agent API that combines STT, TTS, and LLM inference into a single endpoint for building conversational voice agents. Founded in 2015 and headquartered in San Francisco, Deepgram built its own end-to-end deep learning models rather than relying on traditional speech recognition pipelines. The result is significantly lower latency and better accuracy than legacy providers, particularly on noisy audio and accented speech. In February 2026, IBM named Deepgram as its first voice AI partner, integrating Deepgram's APIs into IBM's enterprise AI stack.

Deepgram's Nova-3 model consistently ranks first or second in word error rate benchmarks across English audio types. On typical business audio (meetings, phone calls, podcasts), Nova-3 outperforms Google Speech-to-Text v2, AWS Transcribe, and OpenAI Whisper on both accuracy and latency. The latency story is what separates Deepgram for real-time applications. Sub-300ms round-trip latency makes it viable for live conversation, whereas many competing services introduce delays that make voice agents feel unresponsive. For synchronous voice agent use cases, this is the most important technical differentiator.

The free tier provides $200 in API credits, which translates to roughly 46 hours of Nova-3 transcription or 45 minutes of Voice Agent API usage. Pay-as-you-go rates start at $0.0043 per minute for Nova-3 speech-to-text and $0.015 per 1000 characters for text-to-speech. The Voice Agent API is priced at $4.50 per hour and bundles STT, TTS, and LLM inference together. For teams building voice agents, this simplifies pricing to a single per-conversation cost rather than managing three separate API bills. Volume discounts are negotiable for enterprise workloads above a certain monthly spend.