Deepgram

Deepgram's multilingual edition of Flux conversational STT, released April 2026, covering English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch in one streaming model. Each turn event lists the detected languages, and hints can change mid-stream.

API model id for Flux Multilingual, which extends Flux's turn-aware streaming recognition to 10 languages including Hindi, Japanese and Russian. An optional language_hint biases output; without one it auto-detects and follows mid-sentence code-switching.

Flux is Deepgram's conversational speech recognition model for voice agents, launched October 2025, with turn detection inside the recognizer via StartOfTurn, EagerEndOfTurn and EndOfTurn events. Streaming-only on /v2/listen; this slug resolves to English flux-general-en.

The exact API model id for English Flux STT, requested as model=flux-general-en on wss://api.deepgram.com/v2/listen. It reports end of turn itself, replacing separate VAD and endpointing; Deepgram reports accuracy nearly identical to Nova-3 streaming.

A provider bundle: Deepgram's Voice Agent API set to Nova-3 for listening, OpenAI GPT-4o mini for reasoning and the Aura-2 Thalia English voice. One WebSocket handles turn-taking, barge-in and function calls; the API reached general availability in June 2025.

Second-generation Aura text-to-speech with 90 voices in English, Spanish, Dutch, German, French, Italian and Japanese, over REST and WebSocket with speed and pronunciation controls. Deepgram now steers English builds to Flux TTS and keeps Aura-2 for the other six languages.

Clinical-audio variant of Nova-3, released March 2025 for batch and streaming, which gained multilingual medical transcription in 10 languages in October 2026. Deepgram reported a 3.44% median WER at launch; nova-3-pharma covers drug terminology separately.

Deepgram's current flagship transcription model, launched February 2025 for batch and streaming audio in 50+ languages, with multilingual code-switching and keyterm prompting. At launch Deepgram reported median WER of 5.26% batch and 6.84% streaming.

Air traffic control variant of Nova-2, released August 2024 for noisy, jargon-heavy radio exchanges between pilots and controllers. English-only, it transcribes both pre-recorded and streaming audio, and Deepgram has not shipped a Nova-3 equivalent.

Deepgram's first-generation text-to-speech model, generally available since March 2024 with 12 English voices named aura-{voice}-en. Followed by Aura-2 and Flux TTS, it stays on sale at half Aura-2's per-character price.

Nova-2 is Deepgram's previous-generation transcription model, generally available since November 2023 and superseded by Nova-3. It covers batch and streaming audio in more than 30 languages plus Spanish-English code-switching, and is still sold as a cheaper option.

A Nova-2 model option whose vocabulary is tuned toward automotive terminology, limited to English (en, en-US). It dates from the Nova-2 generation of late 2023, and the current Nova-3 line offers no automotive-specific version.

Tuned for a person talking to an automated system such as an IVR, this Nova-2 option transcribes English only. For new voice-agent work Deepgram now points developers to Flux, which adds turn detection to recognition.

Drive-thru ordering audio is the target of this English-only Nova-2 variant, called as model=nova-2-drivethru. It belongs to Deepgram's previous model generation, and Nova-3 has no drive-thru counterpart.

Built for earnings-call style recordings with several speakers and uneven audio quality, nova-2-finance is an English-only Nova-2 domain model. Nova-3 has no finance variant but accepts keyterm prompting for domain terms.

Nova-2's medical-vocabulary model for English clinical audio, retrained in May 2024 for a 16% relative gain in key-term recall, per Deepgram. Superseded by nova-3-medical, which Deepgram reports lifts keyword recall another 10.6%.

Meant for conference rooms where several people share one microphone, nova-2-meeting is an English-only option in the Nova-2 family. Deepgram keeps Nova-2 models for existing deployments and points new projects to Nova-3 or Flux.

English-only Nova-2 model for low-bandwidth telephone audio, and the base from which Deepgram derived its voicemail variant. Part of the previous generation; Nova-3 is Deepgram's recommendation for new transcription projects.

This Nova-2 variant is optimized for speech in video soundtracks and supports English (en, en-US) only. Its generation has been superseded by Nova-3, which has no video-specific version.

Derived from Deepgram's phonecall model, this English-only Nova-2 variant handles short, low-bandwidth clips with a single speaker, such as voicemail messages. It remains available but predates the current Nova-3 generation.