Expressivity

The expressivity parameter shifts a Flux TTS voice’s delivery register along a calm to animated axis. Negative values typically produce calmer, steadier, slightly lower-pitched delivery; positive values typically produce more animated delivery with a wider pitch range.

Every Flux voice is tuned by Deepgram to its most natural delivery out of the box, and 0 is that tuned default. Think of expressivity as a signed offset from the default rather than a required setting: the voice already sounds good at 0, and the parameter shifts its register when your use case calls for it.

Expressivity is a register dial, not a quality slider. The calm end is a deliberate production choice for contexts like support and IVR, not a lesser mode; the animated end is a different character, not a better one.

Parameter reference

Values must be whole numbers, and omitting the parameter is equivalent to 0. The server rejects invalid values rather than clamping or rounding them: out-of-range values return EXPRESSIVITY_OUT_OF_RANGE, and fractional values such as 1.5 return EXPRESSIVITY_INCREMENT_INVALID.

What each value sounds like

Because each voice has its own natural character, the same value can sound different across voices: a 2 may be a subtle change on one voice and a pronounced one on another.

Matching register to use case

These are starting points, not rules. The right value depends on the voice, the content, and the audience, so audition candidate values with your own prompts before committing to one.

Choosing a value

The fastest way to choose is to hear a few values side by side before hardcoding one. Generate the same line at -1, 0, and 1 with the requests in the Usage section below. For a first listen, talk.deepgram.com offers the full -2 to 2 range, labeled Calm to Animated, on a subset of voices; to audition your exact script, voice, and encoding, use the API. Expressivity controls in the Playground are coming soon.

Because the parameter is in beta, re-validate your chosen value when Deepgram ships model updates rather than treating it as a fixed contract.

Usage

Set expressivity as a query parameter on either /v2/speak transport, or as an agent.speak.provider field in the Voice Agent API. On streaming, the value is fixed when you open the connection and applies for its duration; on batch, it applies to that request.

Streaming

wss://api.deepgram.com/v2/speak?model=flux-haley-en&encoding=linear16&sample_rate=24000&expressivity=-1

Batch

$
curl "https://api.deepgram.com/v2/speak?model=flux-haley-en&expressivity=2" \
>
     -H "Authorization: Token YOUR_DEEPGRAM_API_KEY" \
>
     -H "Content-Type: application/json" \
>
     -d '{"text": "Your appointment is confirmed for 3pm tomorrow."}' \
>
     --output audio.mp3

Voice Agent

In the Voice Agent API, set agent.speak.provider.expressivity in the Settings message rather than a query parameter. It takes the same -2 to 2 values and applies for the whole session.

1
{
2
  "agent": {
3
    "speak": {
4
      "provider": {
5
        "type": "deepgram",
6
        "version": "v2",
7
        "model": "flux-haley-en",
8
        "expressivity": -1
9
      }
10
    }
11
  }
12
}

See Configure the Voice Agent.

Behavior notes

• The register shift is intentional behavior. Moving away from 0 changes pitch range, pacing variation, overall pitch, and timbre (the voice’s tonal color) together, so the voice’s character audibly changes toward the ends of the range. This is the parameter working as designed, not a side effect.
• It is not a speed or pitch-shift control. Negative values do not slow the voice down, and positive values do not speed it up. For speaking rate, use speed (on Flux TTS, 0.85 to 1.15 in 0.05 increments).
• Animated delivery naturally varies more. At higher values, delivery tends to vary more between generations of the same text; that variation is part of what makes speech sound animated and human. Deepgram’s output-consistency characteristics for Flux TTS are measured at the default (expressivity omitted or 0).
• Effect and risk grow with magnitude. The further from 0, the stronger the shift, and the higher the chance of hallucinations or pronunciation errors. The range is kept narrow during beta for this reason; audition any non-default value before production.
• Comparing text-to-speech providers? Evaluate Flux TTS at the default (0) or above. The calm end deliberately narrows delivery for production contexts like support and IVR, so it does not showcase the model’s expressive range.

FAQ

Does this change which voice I’m using?

You are always getting the same underlying voice. That said, expressivity does more than adjust energy: pitch and timbre shift too, so the voice’s character changes noticeably toward the ends of the range, the way a person sounds different reading a bedtime story versus hosting a game show.

Can I change it mid-conversation?

No. expressivity is set when you open the connection and applies for the whole session.

How does it interact with speed?

They are independent controls: speed adjusts speaking rate, and expressivity adjusts delivery register. They can be combined, but larger values of both together are more likely to sound less natural, so audition combined settings.

Why does the same value sound different on different voices?

Each voice has its own character and its own sensitivity to the dial. 0 is every voice’s tuned default, but a given step away from 0 lands differently from one voice to the next, so the same value can be a subtle shift on one voice and a pronounced one on another. Audition your specific voice.