Delivery is a request parameter.

Most engines give you a voice and a volume knob. This one exposes the prosody itself: intensity, melodic range, pacing, speed and gain, each a continuous field on the ordinary request. The figures on this page are what those fields actually did when we measured them.

7.9

semitones of pitch range between the dial’s ends

The dial that decides how a line lands

temperature sets prosodic variation — pitch range and melody — from 0.1 to 1.2. It is the control a listener actually perceives as the difference between a line recited and a line performed, and unlike a preset it is continuous.

Measured: one voice, one sentence, three seeds at each setting, rendered on the production API on 2026-07-21. Pitch range read from an autocorrelation pitch track as the p5–p95 spread in semitones. At 0.3 the read holds a 10.08-semitone range; at 0.9 it opens to 17.99 — a 7.91-semitone difference in melodic range from one field, on the same voice saying the same words. The curve is not monotonic: it peaks near 0.9 and comes back to 13.30 at the 1.2 ceiling, which is the kind of thing you only learn by measuring rather than by reading a range in a doc.

measured 2026-07-21 · gandr-mia · 3 seeds per setting

temperature   pitch range (p5-p95, semitones)
0.1                  10.65
0.3                  10.08   <- flattest
0.5                  10.60
0.7                  14.73
0.9                  17.99   <- widest
1.2                  13.30

# same voice, same sentence, expressiveness 0.9, cfg_weight 0.3

The rest of the surface

The dials interact, and the interaction is the part worth knowing. Intensity alone, driven to its maximum against tight guidance, measurably FLATTENS a read: we measured 10.93 semitones of range at maximum intensity against 13.69 on a lower intensity paired with loosened guidance, three seeds each. Raise expressiveness and lower cfg_weight together, or the bigger delivery only gets hurried.

expressiveness — emotional intensity, 0.25 flat to 2.0 animated. Default 0.5.
cfg_weight — guidance strength, which also sets pacing: 0.2 slow and spacious to 1.0 tight and brisk.
speed — 0.6 to 1.5, applied after synthesis, so wording and voice are untouched.
volume — 0.5 to 2.0, applied after mastering with a soft ceiling, so it never hard-clips.
seed — an integer that makes a render reproducible: the same seed, text, voice and parameters return the same audio, which is what lets you regression-test a prompt change without the voice moving underneath you.

The emotion field, and what it actually is

A body field names an emotional read — calm, apologetic, excited, curious and others — and sets tuned dial values under the hood. Explicit numeric fields always override it.

It is a shortcut over the dials, not a separate capability, and we measured how many distinct reads sit behind the names: sixteen accepted names resolve to eight distinct presets, with several names sharing one (bored, calm and tired return byte-identical audio at a fixed seed, as do angry and surprised). Their measured pitch range spans 9.94 to 13.72 semitones — real separation, but a narrower span than the dials reach on their own. We publish that because it is what the measurement says: reach for the presets when you want a sensible read in one field, and for the dials when you want a specific one.

the eight distinct presets behind sixteen names

neutral      = sending no emotion field at all
calm         = bored = tired
angry        = surprised
confident    = curious = happy
frustrated   = scared
apologetic   = content
anxious      · excited · sad     (distinct)

# fixed seed, one voice, one sentence; grouped by identical
# output. 49 takes, measurements.csv published alongside.

And the detail you place by hand

These ride inline on every surface — bytes, SSE and WebSocket alike — and none of them costs a round trip or a preprocessing step. They are read at synthesis time, and first audio holds its published 107 ms p50 with every field on this page set.

<spell>TKT4829XB</spell> — reads character by character, with letter and digit groups paced naturally. For confirmation codes, order IDs and serials.
<break time="800ms"/> — a real pause, rendered as a natural beat, exactly where the read needs air. The tag is never spoken.
pronunciation_dict — per-request sounds-like replacements for proper nouns and domain terms, so a brand name is right on the first call, not after a support ticket.

Notes — an engineer's checklist

01How much can the delivery actually change?

Measured on one voice and one sentence, three seeds per setting: the prosody dial moves pitch range from 10.08 semitones at its flattest to 17.99 at its widest, a 7.91-semitone spread. That is the melodic difference between a recited line and a performed one, from a single request field. The full table and the audio are published on this page.

02Is expression a separate model or a slower path?

Neither. Every control here is a field on the ordinary synthesis request, conditioning the same voice on the same endpoint. There is no expressive tier and no second model to route to — first audio holds its published 107 ms p50 with the fields set.

03How many distinct emotional presets are there?

Sixteen names are accepted and they resolve to eight distinct presets — several names share one, which we verified by rendering all sixteen at a fixed seed and grouping byte-identical output. The presets are a convenience over the numeric dials; the dials themselves are continuous and reach a wider range.

04Can I reproduce an exact take?

Yes. Send a seed: the same seed, text, voice and parameters return the same audio, per serving region. Omit it for a fresh natural take on every call. That is what lets you diff a prompt change without the voice moving underneath you.

05Do these controls work on a cloned voice?

Yes, on ready-made and cloned voices alike, and across all 23 languages. The reference clip supplies the identity; these fields direct the performance of it.

06What sample rate does the audio come back at?

24 kHz mono PCM by default. Set output_sample_rate if your stack wants something else — 8 kHz for a telephony trunk, for instance — so no transcode sits in the hot path. The default is the full-rate audio, not the telephony rate.