Deepgram

What is Deepgram?

Deepgram is an AI speech and voice platform for developers. Its APIs process live or recorded audio, return transcripts, generate speech from text and support real-time conversational agents. The platform can be used through hosted APIs and SDKs, with regional, dedicated and self-hosted options available for organizations that need more control over infrastructure or data location.

The main reason to use Deepgram is to add voice functionality to an existing product. Typical examples include call transcription, live captions, contact-center analytics, speech-enabled interfaces and voice agents that combine speech recognition with an LLM and text-to-speech.

What Deepgram actually does

Speech-to-text for live and recorded audio

Deepgram's core capability is automatic speech recognition. Developers can send streaming audio over a WebSocket connection or process pre-recorded audio through an API. The result can be used for captions, searchable transcripts, call records, meeting or interview transcripts and downstream analysis. Models such as Nova-3 and Flux are aimed at different real-time and speech-processing workflows, while Whisper Large Cloud is also available.

Text-to-speech

Deepgram converts text into streamed or file-based audio through its Aura and Flux text-to-speech models. This is useful for applications that need spoken responses, such as phone systems, accessibility features, voice interfaces and conversational agents. It is an API capability rather than a consumer voice-creation studio, and the quality, latency and language options depend on the selected model and deployment.

Voice agents

The Voice Agent API provides a real-time workflow that combines listening, language-model reasoning and speaking. Developers can use Deepgram's speech services with Deepgram-managed LLM options or configure customer-selected providers and endpoints, including OpenAI, Anthropic, Google, NVIDIA, Groq and Amazon Bedrock configurations. This lets a product handle spoken conversations without requiring the development team to assemble every audio component independently.

Audio intelligence

Deepgram also supports analysis of conversations and transcripts, including summarization, topic detection, sentiment analysis and intent recognition. These features are relevant to contact centers, call analytics and quality workflows, but they are still delivered as developer services rather than as a complete end-user customer-support application.

How people use Deepgram

A typical implementation starts with an account and API key. The developer selects a speech, voice or intelligence endpoint, sends audio or text through the REST API, WebSocket API or an official SDK, and then routes the returned transcript, audio or analysis into the application's database, interface or workflow.

• Contact centers: Transcribe calls, generate summaries and identify topics, sentiment or intent.
• Voice applications: Build spoken interfaces and conversational customer-support agents.
• Media and content workflows: Transcribe interviews, podcasts and other recorded audio.
• Accessibility and captions: Produce real-time or recorded speech captions.
• Telephony: Connect speech recognition and synthesis to systems such as Twilio, Amazon Connect or other supported voice infrastructure.
• Private deployments: Use regional endpoints, dedicated infrastructure or self-hosted services where deployment control matters.

Deepgram provides official SDKs for Python, JavaScript/TypeScript, Go, .NET, Java and Rust, along with command-line and browser-based developer tools. Integrations and reference architectures include Twilio, LiveKit, Pipecat, AWS, Amazon Connect, Amazon Bedrock, Amazon SageMaker, Five9 and Pipedream.

Pricing and access

Deepgram uses usage-based pricing rather than a conventional unlimited subscription. New accounts receive $200 in initial usage credit, but this is introductory credit rather than a continuing free plan. The lowest listed standard Pay As You Go rate is $0.0043 per minute for pre-recorded Nova-3 monolingual speech-to-text. Other services are priced separately, including streaming transcription, text-to-speech by character, Voice Agent usage by minute and Audio Intelligence usage.

Limits vary by service. Current documentation lists Pay As You Go concurrency limits of up to 50 REST speech-to-text connections, 150 WebSocket speech-to-text connections, 45 text-to-speech connections, 45 Voice Agent connections and 10 Audio Intelligence connections, with higher limits available through Growth or Enterprise arrangements. Growth plans require an annual commitment starting at $4,000 per year, while Enterprise pricing is custom.

Privacy and deployment considerations

Deepgram's data handling depends on the endpoint, account configuration and deployment. Its documentation states that API requests participate in the Model Improvement Program by default and may be retained for model improvement. Customers can set mip_opt_out=true; opted-out requests are not retained after processing and are not used for model improvement, although request metadata and usage logs may still be recorded.

Regional endpoints can support processing and storage requirements involving regions such as the European Union, Australia or India when configured appropriately. Deepgram states that data is encrypted in transit and at rest and documents support for SOC 2, HIPAA, PCI DSS, GDPR, CCPA and Australian Privacy Principles. Organizations should still review the applicable retention settings, contractual terms and any separate policies of third-party LLM providers used in Voice Agent workflows.

Who Deepgram is for

Deepgram is a strong fit for software developers, startups, product teams, contact-center companies and enterprises that need programmable voice capabilities. It is particularly relevant when an organization wants control over the surrounding application, model selection, integrations or deployment environment.

It is less suitable for a nontechnical user who simply wants to upload recordings and edit transcripts in a finished workspace. In that case, a dedicated tool such as Otter may be a more direct fit. Deepgram can provide the underlying transcription and voice infrastructure for such workflows, but most production use requires API integration, monitoring and application development.

Key strengths and limitations

• Strengths: Broad coverage across transcription, speech synthesis, audio analysis and voice agents; streaming support; multiple SDKs; customer-selected LLM options for voice workflows; usage-based access; and regional, dedicated or self-hosted deployment choices.
• Limitations: It is primarily infrastructure rather than a polished consumer application; pricing varies by service and model; usage and concurrency limits apply; production setup requires engineering work; and default model-improvement participation requires an explicit opt-out for requests that should not be retained for that purpose.

Deepgram is best understood as a voice AI building block. Compared with a general-purpose chatbot or a standalone transcription app, it offers more control over how speech services are embedded into a product, but demands more technical work from the organization using it.