Deepgram
AI speech-to-text API built for developers with industry-leading speed and accuracy.
Quick take
Deepgram is the best speech-to-text API for most developers. The combination of accuracy (4.7% WER), speed (sub-300ms streaming), and price ($0.26/hour) is hard to beat. AssemblyAI is the closest competitor, with slightly lower accuracy but stronger LLM-powered features (summarization, topic detection via LeMUR). Choose Deepgram for raw transcription speed and cost. Choose AssemblyAI for post-processing intelligence.
Overview
Deepgram is a speech-to-text API built for developers who need fast, accurate transcription at scale. Unlike consumer tools (Otter, Fathom), Deepgram is pure infrastructure: you send audio in, you get text back. The company has raised $86M, employs 101-250 people, and processes billions of minutes of audio per year. With 440 G2 reviews at 4.6/5 and 76% satisfaction, Deepgram is one of the most adopted speech APIs in the developer community. Their Nova-2 model consistently ranks at or near the top of independent accuracy benchmarks.
Key strengths
Speed and accuracy are the dual advantages. Deepgram returns streaming transcription results in under 300 milliseconds, fast enough for real-time use cases like live captions, voice agents, and in-call assistants. Nova-2 achieves approximately 4.7% Word Error Rate on meeting-style audio, matching or beating Google, Azure, and OpenAI Whisper. Pricing is aggressively low at $0.0043 per minute ($0.26/hour), undercutting most competitors. The API supports 30+ languages, custom vocabulary, speaker diarization, and topic detection.
Limitations
Deepgram is not a meeting recording tool. It does not join meetings, capture audio, or manage bots. You need to bring your own audio from another source (Recall.ai, a browser extension, a phone system). The free tier gives $200 in credit, which sounds generous but goes fast at production scale. Enterprise features (on-premise deployment, custom models) require custom pricing conversations. Documentation is good but the ecosystem of tutorials and community content is smaller than AWS Transcribe or Google Speech.
Pricing breakdown
Pay As You Go: $0.0043/minute for Nova-2 (pre-recorded), $0.0059/minute for streaming. $200 free credit to start. Growth and Enterprise tiers offer volume discounts, custom models, and on-premise deployment. No minimum commitment on Pay As You Go.
Who should use Deepgram
Developers building voice applications, transcription features, or real-time captioning. Teams that already have an audio capture pipeline and need the best accuracy-to-price ratio for the transcription step. If you need an all-in-one meeting recording solution (capture + transcription + summaries), look at Recall.ai + Deepgram as a stack, not Deepgram alone.
Verdict
Deepgram is the best speech-to-text API for most developers. The combination of accuracy (4.7% WER), speed (sub-300ms streaming), and price ($0.26/hour) is hard to beat. AssemblyAI is the closest competitor, with slightly lower accuracy but stronger LLM-powered features (summarization, topic detection via LeMUR). Choose Deepgram for raw transcription speed and cost. Choose AssemblyAI for post-processing intelligence.
Key features
- Real-time streaming transcription
- Pre-recorded audio API
- Custom model training
- Speaker diarization
- Language detection
Pros and cons
Pros
- + Fastest streaming transcription available
- + Competitive accuracy
- + Good developer experience and docs
Cons
- - Fewer features than AssemblyAI for intelligence
- - Custom models require enterprise plan
- - Startup, long-term viability uncertain
What users say
Works well with different voices and accents, even with background noise or strong accents.
G2
Can sometimes struggle when the audio is very noisy or when multiple people speak over each other.
G2
