Deepgram: Aura-2

Aura-2 is Deepgram’s text-to-speech family with voices across seven languages: English, Spanish, German, French, Dutch, Italian and Japanese. It converts the supplied text into speech for customer-service replies, telephone menus, explanations and narration. The manufacturer describes pacing, expressiveness and conversational fillers as responses to the wording and context of the supplied text.

Match the voice to the language and task

Each voice is identified by its family, speaker and language. The catalogue distinguishes accents and delivery characteristics rather than treating every speaker as interchangeable. Thalia is an energetic American English option; Draco is a British voice associated with storytelling; Celeste is a Colombian Spanish voice associated with conversational and announcement uses. Choose the voice with the actual script and listener in mind.

Selected Spanish voices support English–Spanish code switching: Aquila, Carina, Diana, Javier and Selena. That capability belongs to those named speakers, not every voice in the family. For a mixed-language passage, preserve the intended language of product names and quoted phrases, and listen to the transition with a speaker whose documented profile supports the task.

Prepare text for speech

Aura-2’s formatting guide recommends sentence-ending punctuation, question marks for questions, commas for natural pauses and short standalone phrases. Give the final spoken text, not an instruction asking the synthesizer to write the message. A useful orientation script is: “Welcome. Choose a language, then select the topic you need. You can repeat the instructions at any time.” The punctuation separates the listener’s actions and gives the model a clearer rhythm.

Spell an identifier as separate letters when it must be read character by character. Define acronyms and avoid a long run-on sentence when several conditions must be explained. The maker also recommends quoting command words and checking a passage aloud before synthesis. For educational or product narration, retain the important terms while turning written headings and dense lists into complete spoken phrases.

Pace and pronunciation

The manufacturer documents speed and IPA pronunciation controls for English and Spanish. Aura-2’s speech-rate range is 0.7 to 1.5; for Spanish, it recommends 0.9 to 1.5 because slower settings may introduce disfluencies. Pronunciation and speed can be combined in this family. IPA overrides can target a specialist name or term, and the Aura-2 convention places the stress marker directly before the vowel.

The pronunciation-control specification allows up to five hundred overrides, with each IPA string no longer than 128 characters, and documents a 2,000-character input-text limit for that workflow. Aura-2 does not support the explicit pause-control feature documented for Flux TTS. Use sentence structure and punctuation to prepare the rhythm, and divide longer material into coherent passages when the chosen workflow requires it.

Audio delivery and use

The manufacturer provides complete-request synthesis and streaming text-to-audio workflows. Its media guide distinguishes raw streaming PCM, mu-law and A-law from file-oriented output options such as WAV, MP3, Opus, FLAC and AAC. The documented PCM/WAV default for the Aura request workflow is 24 kHz. File containers, sample rates and encodings form valid combinations; a telephone connection and a downloadable narration file can require different choices.

Aura-2 remains a separate family from first-generation Aura and from Flux TTS. Flux’s cross-turn state, interrupt feedback and batch-only pause controls are not features to infer from an Aura-2 voice name. The useful output is a spoken version of the actual script in a suitable language, accent and audio representation. Review names, language switches, punctuation-led pacing and important instructions with the intended listener before reusing the recording.