Deepgram review: Speech-to-text, text-to-speech and voice agent APIs for developers, billed by usage

What Deepgram does

Deepgram is a voice AI platform for developers and businesses that need to add speech to their products. It offers three main building blocks: speech-to-text for live or pre-recorded audio, text-to-speech, and a Voice Agent API that joins transcription, a language model and spoken replies into a single streaming connection. The speech-to-text side includes Flux STT and Nova-3 models, with support for 60+ languages, and optional extras such as redaction, keyterm prompting, entity detection and speaker diarization. Audio intelligence features cover summarization, topic detection, sentiment analysis and intent recognition.

Pricing is usage-based rather than per seat. The Pay as you go plan starts with a free $200 credit, has no minimums or expiry on that credit, and needs no credit card to begin. Growth is a pre-paid annual option starting at $4,000 a year, with savings of up to 20% and higher concurrency. Enterprise is custom priced and adds options such as self-hosted deployment in a private VPC, on-premises hardware or an air-gapped setup. Billing units differ by product: per minute for speech-to-text, per 1,000 characters for text-to-speech, and per minute of connection time for Voice Agent. For example, Flux STT English is listed at $0.0077 per minute on Pay as you go and $0.0065 per minute on Growth. Stated limits include 150 concurrent streaming transcription connections on Pay as you go and 225 on Growth.

• Real-time and pre-recorded speech-to-text
• Flux STT and Nova-3 models
• 60+ languages for transcription
• Text-to-speech with Flux TTS, Aura-2 and Aura-1
• Voice Agent API with turn detection and function calling
• Audio intelligence: summaries, sentiment, intent, topics
• Keyterm prompting for custom vocabulary
• Cloud, VPC, on-premises or air-gapped deployment

Our take

Deepgram suits developers and product teams building voice features such as live transcription, call analysis or voice agents, and who are comfortable paying for what they use. The free $200 credit with no card lets you test before committing, and one vendor covers speech-to-text, text-to-speech and agent orchestration. Teams that need data to stay in their own environment can look at the Enterprise self-hosted option. The trade-offs: pricing is split across many products and models, so estimating a monthly bill takes some work. Growth needs a yearly commitment starting at $4,000, with overages charged at a 10% premium. The pages list no consumer apps, so this is not aimed at people who just want a transcription app. Self-hosting needs NVIDIA GPUs and an Enterprise agreement.