Deepgram

Pros and cons

About Deepgram

Deepgram is an enterprise voice AI platform that offers speech-to-text, text-to-speech, and a Voice Agent framework to tie its STT and TTS services with third-party LLM APIs. It also offers increasingly pertinent features such as diarization, advanced turn-taking detection, and PII redaction as add-ons, helping companies meet compliance requirements.

Due to its end-to-end deep learning structure, Deepgram's STT is able to achieve high accuracy and low latency, which is ideal for real-time transcriptions. This, along with its ease of use out of the box and custom vocabulary tools, has earned it praise from developers and enterprises.

Deepgram is also considered a feature-rich platform with inclusions such as smart formatting, summarization, and topic detection.

The STT API itself offers smooth integration and flexible deployment options that include Virtual Private Cloud, cloud, and on-premise options.

It is because of these details that Deepgram is ideal for AI and voice agents (especially real-time chatbot responses), customer support and sales, and teams that use domain-specific jargon.

Users do, however, share some criticisms of Deepgram's STT. One such complaint is that their language support is rather limited.

Their English, Spanish, and Portuguese transcriptions are on point, and their platform supports 40+ languages (fewer than other platforms). However, there's a noticeable reduction in accuracy for those less common languages.

An additional complaint is that transcription accuracy drops in edge cases. They might require some additional cleanup, especially when the audio includes heavy accents.

Below, we do a deeper dive into Deepgram's accuracy, features, pricing, and more.

Available Deepgram models

Deepgram offers two foundational model families for their speech-to-text API — Nova-2 and Nova-3 — alongside industry-tuned variants and a conversational model for voice agents. Most users are impressed with their speed, along with their low WER compared to other transcription models evaluated.

• Nova-3

  Flagship

  Deepgram's newest general-purpose model, touted for speed and the ability to scale. Optimized for accuracy, multi-speaker environments, and noise robustness.

  • 5.26% reported WER (batch)
  • 45+ languages
  • Streaming + batch
  • Easy fine-tuning for developers
  • Multilingual codeswitching

• Nova-2

  Legacy

  A long-standing ASR model known for its speed and low WER relative to competing transcription models.

  • Streaming + batch
  • Fine-tuning support

• Flux

  Voice agents

  A specialized, low-latency conversational model engineered for real-time voice agents — tracks conversational flow and handles dynamic turn-taking.

  • Sub-second latency
  • Turn-taking detection

• Industry-tuned models

  Vertical

  Nova variants trained on benchmarks and terminology tailored to specific industries. So far covering finance, meetings, and call centers/customer support.

  • Finance
  • Meetings
  • Call centers
  • Keyword boosting for proprietary terms
  • Custom industry word lists

For those industries (and agents) that aren't covered, users are able to use keyword boosting to inject proprietary terms and brand names or submit a list of industry words and phrases.

Deepgram core transcription and augmentation features

Deepgram is considered feature-rich by many of its users, including the ability to enable numeral conversion to convert numbers to digits and detect filler words. Deepgram also offers:

• Speaker diarizationIdentifies and labels each speaker's words.
• Smart formattingCapitalization, paragraphing, numbers as digits, and punctuation.
• Filler word detectionDetects disfluencies such as “um”, “uh,” and “hmm”.
• Vocabulary weightingImproves critical word recall rates.
• RedactionRemoves sensitive and personal information.
• Multilingual supportTranscribes speakers from different countries.
• SummarizationCaptures the most vital details of a call in a paragraph.

It's important to note that some of these features, such as speaker diarization and redaction, are only available as add-ons, with an additional charge of $0.12/hour. Additionally, there are restrictions on languages and which models are supported for streaming.

Keyterm weighting also comes at an additional charge of $0.08/hour.

Considering the role these features now play in the workflows of enterprise productions, it's vital to consider these costs in your total cost of ownership.

Processing modes

Real-time streaming

High-throughput WebSocket

Batch / async

REST API endpoints

Deepgram STT deployment and infrastructure

Deepgram offers a number of different deployment options including Cloud, VPC, dedicated cloud services, and bare-metal server options to meet different customer needs and requirements:

• Managed Cloud: multi-region endpoints that don't require any infrastructure maintenance. Supported via their standard API in both the “Pay-As-You-Go” and “Growth” plans.
• Self-hosted/on-premise: Deepgram uses Docker and Kubernetes for containerized deployment, enabling virtual private clouds (VPCs) or an air-gapped data center to meet HIPAA and SOC 2 regulatory framework requirements.

Core accuracy benchmarks for Deepgram

Deepgram reports that their Nova-3 model batch delivers its lowest error rate of 5.26% on a generalized audio set. This means their STT displays roughly 5 errors out of every 100 words in a transcript.

Top enterprise STT models typically achieve Word Error Rates between 5 and 7%, or 93–95% accuracy, on clean laboratory benchmarks, with clear English spoken and without background noise.

While accuracy will change depending on the production environment, telephony conditions, and language spoken (most models are tested on English data sets), you can use these numbers as a baseline to understand the potential accuracy you'll receive if you choose Deepgram as your STT provider.

Deepgram's tiered pricing and consumption models

Deepgram uses pre-purchased credits, which are consumed by millisecond usage (based on the total time a WebSocket is open from endpoint to endpoint).

Their tiered pricing starts with their pay-as-you-go plan, which grants a $200 one-time credit upon opening an account. From then on, users pre-pay for credits at a rate of $0.46/hour for their monolingual Nova-3. Multilingual batch and streaming can cost up to $0.55/hour.

Their growth plan allows you to pre-pay credits for use throughout the entire year, starting with a minimum of $4k in credits, equivalent to ~10,250 hours, at their rate of $0.39/hour. Their multilingual model runs for $0.47/hour.

They offer add-on features that are essential to most enterprise productions, including speaker diarization and PII redaction for an additional $0.12/hour on the pay-as-you-go plan and $0.10/hour on the growth plan.

No matter which pricing tier you choose, you will receive these basics:

Core processing and architecture

• Asynchronous (batch) processing: transcribe pre-recorded audio or video files via standard REST API endpoints.
• Synchronous (real-time streaming) processing: transcribe live audio feeds with sub-second latency using WebSocket (WSS) connections.
• Multichannel support: transcribe multi-channel audio files, processing each channel independently while maintaining accurate word timings.
• Broad audio format support: natively supports standard formats, encodings, and sample rates (including MP3, WAV, FLAC, AAC, OGG, AMR, and raw PCM).

Model availability

• Nova-3 engineFlagship general-purpose model, optimized for accuracy, multi-speaker environments, and noise robustness.
• Flux engineLow-latency conversational model engineered for responsive real-time voice agents and chatbots.
• Global language supportAccess to 45+ languages and their localized/regional models.
• Automatic language detectionDetects the dominant language in the audio and routes it to the corresponding model.
• Multilingual codeswitchingTransitions transcription when speakers shift between languages within one conversation.

Smart formatting and readability

• Smart formatting: applies logical punctuation, capitalization, paragraph structure, and entity formatting (dates, phone numbers, currencies).
• Speaker diarization: identifies multiple distinct speakers and labels who spoke what phrase and when.
• Filler word detection: identifies and separates standard vocal disfluencies from core conversational content.
• Profanity filtering: option to automatically detect and sanitize explicit language from output transcripts.

Custom vocabulary and search

• Keyterm weighting: inject up to 100 custom terms at inference time to boost word recognition accuracy.
• Find and replace: search and systematically replace specific words or phrases during transcription generation.

Result processing and security

• Callback integration: direct transcripts to a designated webhook URL upon compilation.
• Tagging and metadata tracking: attach custom identifiers and tags to API payloads for usage tracking, filtering, and logging.
• Data privacy compliance: standardized information security protocols protect data streams across the platform pipeline.

Customer support

Deepgram continues its tiered service for customer support, with the pay-as-you-go tier often receiving community-driven support, while enterprise customers receive contracted enterprise support.

Much of their frontline support toolkit is technical as well as self-service, often relying on developers and engineering teams for smooth integration.

If you are building on a standard pay-as-you-go or free tier, official direct human support is highly limited. Technical troubleshooting for standard tiers relies on public documentation, the GitHub Discussions forum, and the Deepgram Discord server.

No matter your pricing tier, Deepgram offers the following support:

• A support bot that can be installed directly into your workspace, providing documentation, answers to technical questions, and instant code examples.
• Developer forums and community through GitHub Discussions and Discord for technical questions, SDK issues, and feature requests.
• Help center and API documentation, with a dedicated status page and a continuously updated help center.

For organizations with contracted annual or enterprise plans, Deepgram provides dedicated support channels. Enterprise customers receive:

• Direct technical assistance from their engineering and customer success teams.
• Structured SLAs for response times.
• Assistance with high-volume scaling or custom model deployments.
• Auditing of the technical support ecosystem and response time guarantees.

FAQs