Deepgram: Flux TTS

Flux TTS is Deepgram’s text-to-speech family for live conversation and fixed spoken content. Its defining feature is conversational continuity: the streaming model carries prosody across turns, so delivery can retain a consistent tone as an interaction continues. It accepts written text and returns speech audio. Flux TTS and the separately named Flux speech-recognition model perform different parts of a voice conversation.

Select an English voice

The documented Flux catalogue speaks English, with American, British, Irish, Australian, Indian, Singaporean and Filipino accents. Voice descriptions help choose a speaker for customer service, telephone menus, narration or casual conversation. For example, the maker lists Haley as a calm American voice and Jack as a professional British voice. Select a voice around the audience and the kind of wording it will read, then keep it stable through the piece or conversation. Other-language speech belongs to a separately documented model choice.

Live turns and interruptions

In the manufacturer’s streaming workflow, text can arrive as the application composes a response and audio returns incrementally. A turn has its own start and completion events; ending the turn completes the remaining synthesis. Cross-turn context preserves prosody within that session. This is useful for a spoken assistant that first explains a choice and then answers a clarification without changing its vocal character.

An interruption cancels the active turn and returns the spoken portion and the text left over. That distinction helps the application continue from what the listener actually heard. For a practical interaction, prepare a short answer with the important instruction first, followed by the explanation. If the listener interrupts after that first instruction, the next response can address the new question instead of assuming the whole explanation was delivered.

Fixed audio and output formats

Batch synthesis takes a complete block of text and returns its audio in one response. It uses the same voice catalogue as streaming and suits recorded telephone prompts, notifications or narration prepared in advance. Batch requests are independent; streaming turn state and interruption events do not apply to a fixed clip.

The streaming output is raw audio without a file container: 16-bit PCM, mu-law or A-law. The maker documents PCM sample rates of 8, 16, 24, 32, 44.1 and 48 kHz; mu-law and A-law use 8 or 16 kHz. Batch output additionally supports MP3, Opus, FLAC and AAC. Select an encoding and container that the intended player or telephone path can interpret rather than treating a stream of raw samples as an ordinary audio file.

Delivery controls and boundaries

The manufacturer documents a speech-rate range from 0.5 to 1.5 in 0.05 increments. Streaming speed changes apply at a segment boundary. A separate beta expressivity control moves delivery from calm to animated. Ordinary SSML and recognised competitor tags are stripped rather than used as expressive instructions; native Flux controls have their own meaning.

Inline pauses belong to batch synthesis, with durations of 500–3,000 ms in 100-ms increments and at most eight markers per request. Text must separate adjacent pauses, and speed is capped at 1.15 when a pause is present. Pronunciation overrides use IPA and are in Early Access; the maker notes that a requested pronunciation can vary across generations. Pronunciation cannot be combined with a non-default speed or with pause controls. For a script with specialist names, listen to those terms in the selected voice and keep the spoken wording explicit.

Deepgram’s documented streaming session lasts up to one hour and closes after sixty seconds without an inbound message unless kept active by the supported connection mechanism. Those session limits concern the live synthesis workflow. A useful deliverable is an intentional voice, listener-friendly text and audio suited to its destination, with live applications tracking what was actually spoken when a turn is interrupted.