Deepgram

Deepgram is a speech AI platform built for production applications. It provides a speech-to-text API for transcribing audio in real time or from recorded files, a text-to-speech API for generating lifelike voice output, and a voice agent API that combines both into a full conversational pipeline. Its Nova-3 transcription model delivers high accuracy with low latency, which makes it the preferred choice for developers building voice agents, call center tools, meeting intelligence applications, and real-time transcription features.

Where OpenAI Whisper is a model you run yourself on batch audio, Deepgram is a managed API designed for scale, with streaming support, enterprise infrastructure, and pricing based on audio minutes rather than compute time.

What does Deepgram do?

Deepgram converts spoken audio into text (speech-to-text) and text into speech (text-to-speech), with a combined voice agent API for building conversational AI that can hear and respond in real time.

The speech-to-text API handles two modes. Streaming transcription processes live audio from a microphone or telephony source and returns partial and final transcripts with sub-300ms latency, which is fast enough for a voice agent to respond naturally without noticeable delay. Pre-recorded transcription processes audio files (meetings, calls, podcasts) and returns a full transcript with richer features like topic detection, sentiment, and intent recognition.

Post-processing features run on the transcript itself: speaker diarization labels each word with the speaker who said it, smart formatting adds punctuation and capitalizes proper nouns, and custom vocabulary lets you feed in product names or technical terms that the model should handle correctly.

The text-to-speech API (Aura model) produces natural-sounding voice output with multiple voice options and streaming delivery so responses start playing before the full text is generated, reducing perceived latency in conversational applications.

Who uses Deepgram?

AI engineers building voice agents use Deepgram for the transcription layer. A voice agent needs to understand what the user said before it can respond: Deepgram converts the user’s speech into text fast enough that the overall turn-taking feels natural rather than laggy. The Nova-3 + Aura combination covers both the input and output audio sides of the agent.

Call center and customer support teams use Deepgram to transcribe live calls and recordings. Speaker diarization separates agent and customer speech, enabling downstream analysis: did the agent follow the script, how long was the customer on hold, what topics came up, what was the sentiment at the end of the call?

Meeting intelligence products like transcription tools, note-taking apps, and video conferencing integrations use Deepgram’s batch transcription API to process recorded meetings, identify speakers, and extract structured summaries.

Media and podcast platforms use it to generate searchable transcripts from audio content at scale, with accurate punctuation and formatting reducing the amount of manual cleanup needed.

How does Deepgram work?

Deepgram’s transcription pipeline uses end-to-end deep learning models trained on large corpora of audio. Unlike traditional speech recognition systems that chain separate acoustic and language models, Deepgram’s models process audio directly into text in a single pass, which is part of why latency is low.

For streaming transcription, you open a WebSocket connection to the Deepgram API, stream audio bytes from your microphone or telephony source, and receive back transcripts as events. The API returns interim transcripts as the model processes incoming audio, then locks in a final transcript when it detects a pause or end of speech. Your voice agent reads the final transcript, generates a response, sends that text to the TTS API, and plays back the audio.

For pre-recorded transcription, you send an audio file or URL via HTTP and receive a structured JSON response with the full transcript, confidence scores, word-level timestamps, and any post-processing features you requested (diarization, topic detection, sentiment, etc.).

Authentication uses API keys. The API accepts most common audio formats: WAV, MP3, FLAC, MP4, OGG, and others. Client libraries are available for Python, JavaScript, Go, .NET, and Rust.

Deepgram use cases

Voice agent transcription layer. A conversational AI that answers customer calls needs to transcribe what the caller says in real time before the LLM can generate a response. Deepgram’s streaming API provides the transcription, the LLM generates the reply, and Deepgram’s TTS API speaks the response. The latency of each step determines whether the conversation feels natural.

Call center quality assurance. Every inbound customer call is transcribed with diarization enabled. A downstream system tags each call with topics, detects customer sentiment, checks whether the agent followed compliance scripts, and flags calls that need supervisor review. This replaces manual spot-checking with coverage of every call.

Meeting transcription and notes. After a recorded meeting, the audio is sent to Deepgram’s batch API. The transcript is split by speaker, topic-detected, and passed to an LLM to produce a structured summary with action items. The result is a searchable, organized record of what was discussed and decided.

Podcast and media indexing. A media platform transcribes its audio archive to make episodes searchable by content. Deepgram processes large batches of audio files, the transcripts are indexed, and users can search for specific topics or quotes across thousands of hours of content.

Is Deepgram free?

New accounts receive $200 in free credits, which covers roughly 26,000 minutes of Nova-3 streaming transcription. After credits are exhausted, pricing moves to pay-as-you-go: approximately $0.0077 per minute for streaming and $0.0043 per minute for pre-recorded audio using Nova-3. Growth plans start at $4,000 per year and provide discounted rates for higher usage volumes. Enterprise pricing is custom and includes SLA guarantees, dedicated support, and on-premises deployment options.

How to get started with Deepgram

Sign up at deepgram.com to receive the $200 credit and generate an API key. Install the Python or JavaScript SDK and run a transcription against a sample audio file:

from deepgram import DeepgramClient

deepgram = DeepgramClient("YOUR_API_KEY")
response = deepgram.listen.prerecorded.v("1").transcribe_url(
    {"url": "https://dpgr.am/spacewalk.wav"},
    {"model": "nova-3", "diarize": True, "punctuate": True}
)
print(response["results"]["channels"][0]["alternatives"][0]["transcript"])

The Deepgram documentation at deepgram.com covers the streaming WebSocket API, available post-processing features, and the TTS and voice agent APIs for building full conversational pipelines.