Deepgram: Uses, Pricing and Evaluation Guide
What is Deepgram?
Deepgram provides speech recognition, speech generation and voice-agent APIs for audio applications. Compare transcription quality and whole-call costs before choosing a model.
Pricing and access
The official page offers $200 in starting credit followed by usage billing. Growth starts with $4,000 in annual prepaid credits. Rates use different units, including audio time, generated characters and connection time; an STT rate is not the price of an entire conversation.
Who should use it?
Consider it when your application processes recordings or reacts to live speech and you have engineers to handle streaming, errors and authentication. A team that only needs searchable meeting notes should compare a finished assistant first. API flexibility adds responsibility for recording consent, storage and application behavior.
A trial that catches consequential errors
Create a permitted audio set with quiet speech, background noise, overlapping speakers and domain terms. Write a reference transcript first. Score names, negations, amounts and speaker assignments separately from overall word accuracy. A transcript that changes “do not renew” to “renew” can read well while reversing the meaning.
Evaluate the whole conversation
Use a simulated booking request containing an interruption, correction and pause. Measure time to the first useful response and whether the assistant waits for the caller to finish. Keep booking writes disabled until transcript and tool arguments are reviewed. Reconnect during a trial to check that retrying does not duplicate a booking.
Cost worksheet and data boundaries
Estimate audio minutes, generated speech, model calls, telephony and retries separately, then divide by accepted completed tasks. Start with synthetic calls and narrow API permissions. Decide where audio and transcripts are retained and how they are deleted. Publisher benchmarks cannot establish accuracy for your microphone or vocabulary.
Alternatives and selection
Compare AssemblyAI and Speechmatics for recognition, and Cartesia for speech generation and conversational infrastructure. Use one labeled audio set across candidates.
Sources checked October 30, 2026: Official product and pricing information.