Deepgram Base: model profile

Overview

Deepgram describes Base as being built on its end-to-end speech-to-text architecture and says it offers a "solid combination of accuracy and cost effectiveness in some cases." Current documentation places Base under Legacy Models, alongside older Nova and Enhanced lines, and states that Enhanced generally has higher accuracy and better uncommon-word handling than Base. Base remains present in Deepgram's 2026 documentation and rate-limit tables, and current API reference pages still default the generic model parameter to base-general.

In current official materials, Deepgram recommends Flux for real-time voice-agent use cases and Nova-3 for highest-accuracy general transcription, especially for noisy, multilingual, far-field, or crosstalk-heavy audio. Base is a family of task-oriented variants rather than one unified model: base-general (default), base-meeting, base-phonecall, base-voicemail, base-finance, base-conversationalai, and base-video.

Capabilities and features

Base variants and their documented positioning:

Transcript-layer features documented for Base:

The pre-recorded API returns transcript alternatives that include a transcript-level confidence number plus per-word {word, start, end, confidence} objects. Example responses include speaker, speaker_confidence, and punctuated_word fields when diarization and formatting are enabled.

Customization

Base supports classic keywords boosting/suppression. Keyterm prompting is Nova-3 only, and Deepgram's "instant self-serve customization without model retraining" messaging applies to Nova-3, not Base. Deepgram also supports account-linked custom trained models via custom_id; the model-options page says these custom models are available only to Enterprise customers.

Language support

The base-general variant supports the following languages and dialect tags: Chinese (zh, zh-CN, zh-TW), Danish (da), Dutch (nl), English (en, en-US), French (fr, fr-CA), German (de), Hindi (hi, hi-Latn), Indonesian (id), Italian (it), Japanese (ja), Korean (ko), Norwegian (no), Polish (pl), Portuguese (pt, pt-BR, pt-PT), Russian (ru), Spanish (es, es-419, es-LATAM), Swedish (sv), Tamasheq (taq), Turkish (tr), and Ukrainian (uk).

All specialty variants (base-meeting, base-phonecall, base-voicemail, base-finance, base-conversationalai, base-video) support English (en, en-US) only.

Deepgram's language documentation states that its English models are designed to handle global English accents and dialects, but transcript output is normalized to standardized American spelling.

Performance and benchmarks

Deepgram's current public documentation does not publish a current official Base-specific WER or CER figure. Current docs publish detailed accuracy claims for Nova-2 and Nova-3, but not an equivalent single-number benchmark for Base.

Latency and throughput

Deepgram's current guidance states that streaming transcription latency is optimized to 300 ms or less, with a typical breakdown of 150-300 ms transcription latency and 200-500 ms total transcript latency end to end depending on network, buffering, and client-side processing. Deepgram recommends sending audio in 20-100 ms chunks; larger buffers increase built-in delay, while smaller chunks increase overhead.

For pre-recorded jobs, Deepgram's quickstart documentation notes a 2 GB maximum file size and states that requests whose processing exceeds 10 minutes for Nova/Base/Enhanced can return a 504 Gateway Timeout.

Deepgram states its self-hosted batch engine can transcribe 1 hour of audio in under 30 seconds; this is platform-level guidance, not a Base-only figure.

Concurrency limits documented for Base: Pay As You Go allows 50 concurrent pre-recorded and 150 concurrent streaming requests; Growth allows 50 / 225 in North America; Enterprise starts at 200 / 300.

Deployment and integrations

API surface

Base uses the same core STT interfaces as the rest of Deepgram's classic transcription stack.

Authentication supports either Authorization: Token or Authorization: Bearer . Deepgram also documents temporary API tokens with a default TTL of 30 seconds.

For pre-recorded audio, a request can send a JSON body with a remote URI or upload binary audio/video directly. Responses return transcript alternatives, overall confidence, and per-word timing/confidence data. Streaming returns a sequence of WebSocket messages including transcript updates plus SpeechStarted, UtteranceEnd, and metadata events.

Deepgram's official SDK surface includes at least JavaScript/TypeScript and Python, and the wider docs ecosystem also has examples for .NET, Go, Java, Python, and JavaScript in self-hosted/STT guides.

Deployment paths

Deepgram's self-hosted hardware guidance for STT lists a recommended baseline of 1 NVIDIA GPU with compute capability 7.0+, 16 GB VRAM, 4 CPU cores, 32 GB RAM, and 50 GB storage. The self-hosted docs also note that authentication is not built in for self-hosted deployments, so teams place Deepgram behind their own API gateway, reverse proxy, or network controls.

Cloud and pipeline integrations

Documented cloud integration patterns include S3-backed batch pipelines using presigned URLs, SageMaker for managed deployment on AWS, self-hosted cloud/on-prem for private inference, and regional endpoints for residency. Deepgram publishes migration guides from AWS Transcribe, Google Speech-to-Text, and OpenAI Whisper to Deepgram.

For real-time media pipelines, an official Twilio/Deepgram guide identifies 8 kHz raw mu-law as the telephony audio shape Twilio sends, which maps to base-phonecall. Deepgram also documents a LiveKit integration path for agent use cases.

For Azure, the reviewed materials point toward self-hosted deployment on Azure infrastructure; Deepgram's self-hosted hardware guidance includes Azure GPU-instance examples. The reviewed docs do not expose a separate Azure-managed Deepgram product surface.

Privacy and compliance

Deepgram's public pricing/security materials state that the platform is SOC 2 Type 1 and Type 2 certified, HIPAA compliant with BAAs for Enterprise customers handling ePHI, GDPR ready with an EU endpoint (api.eu.deepgram.com), and CCPA compliant.

On retention and privacy mechanics, the reviewed public docs describe data-residency options, a model-improvement opt-out (mip_opt_out), and flexible retention language in trust/security materials, but they do not present one globally applicable default-retention statement for Base.

Pricing

Platform-level plan details on the current public pricing page: Pay As You Go includes a $200 free credit, Growth starts at $4K+/year with pre-paid credits, and Enterprise is custom. Historically, Deepgram's 2023 benchmark whitepaper showed Base price bands materially below one dollar per hour of audio, varying by annual volume; those figures are not the current 2026 public price card.

Development and ownership

Deepgram develops and operates the Base model family as part of its speech-to-text platform. Base is built on Deepgram's end-to-end speech-to-text architecture. Release date, training data, and parameter counts are not publicly disclosed in the reviewed sources.

Release history

Not publicly disclosed. The reviewed sources establish that Base is listed under Legacy Models in Deepgram's 2026 documentation, that it remains operational in the public API surface with base-general as the default model parameter, and that Deepgram's 2023 benchmark whitepaper contained historical Base pricing; the sources do not state original release dates or version history for the Base family.