Voices

When to use it

Sound like one company

One cloned voice across every call, every notification and every recording, instead of a different stock voice per surface.

Sound like a person

Clone a founder, a presenter, a support lead — with their consent — and the product speaks in a voice customers already know.

Speak another language as yourself

A cloned voice carries into live interpretation.

Just need a voice?

The built-in library covers most cases and needs no upload.

The invariant

Every voice resolves to a stored reference clip. This is not a convention — it’s what keeps a voice stable.

The TTS model with no speaker conditioning samples a new random voice on every generate() call. In a batch job you’d notice a strange timbre; in a realtime session you get a caller whose agent changes voice between sentences. So:

Cloned voices store the audio you uploaded.
Built-in voices are designed once through the voice-design vocabulary, and the generated clip is frozen as their reference.

Two rules follow, and they matter if you’re working on this path:

Never send prompt_key without ref_audio_url — the serving backends only read the cache key inside that branch.
For realtime, write the voice to session.audio.output.voice. Nothing else is read.

Clone a voice

Three to ten seconds of clean speech is enough. Longer is not better; clean is better. One speaker, no music, no overlap — see Voice cloning best practices for what that means in a recording booth and why it matters so much. Cloning is a multipart/form-data upload, which the generated JSON client can’t express — create_voice is in the SDK but takes no arguments, so post the form directly:

import httpx

with httpx.Client(base_url="https://api.eesi.ai", timeout=120) as http:
    response = http.post(
        "/v1/voices",
        headers={"Authorization": "Bearer sk-eesi-..."},
        files={"file": ("sarah.wav", open("sarah.wav", "rb"), "audio/wav")},
        data={
            "name": "Sarah",
            # A voiceprint is biometric data. Omit this and the upload is
            # refused with a 400 before the audio is even read.
            "consent_attested": "true",
            # Omit this and the clip is transcribed for you — and the model is
            # conditioned on whatever that transcription said.
            "ref_text": "Thanks for calling — I can help you with that right away.",
        },
    )
    response.raise_for_status()
    print(response.json()["voice_id"])

curl https://api.eesi.ai/v1/voices \
  -H "Authorization: Bearer sk-eesi-..." \
  -F name=Sarah \
  -F consent_attested=true \
  -F file=@sarah.wav

Cloning is asynchronous — the voice appears immediately with a status that moves to ready once the reference clip is transcribed and encoded. Poll client.get_voice(voice_id) or just watch the library.

Built-in voices

Built-in voices are global rows shared across organizations, seeded per environment:

python -m api.scripts.seed_builtin_voices

Idempotent, but it needs a live TTS backend — it synthesizes each voice’s reference clip. Until it runs, the library is empty and realtime falls back to the serving backend’s pinned default. If voices look missing in a fresh environment, this is almost always why.

Design a voice

Rather than cloning, describe one. The accepted vocabulary comes from the backend:

options = client.voice_design_options()
print(options)   # ages, genders, accents, timbres, pacing

Preview

Preview synthesizes one fixed sentence with the voice and streams back mp3 — there is no text to choose, which is what makes two voices comparable. It isn’t in the SDK yet (the SDK decodes JSON, and this returns audio):

import httpx

response = httpx.post(
    f"https://api.eesi.ai/v1/voices/{voice_id}/preview",
    headers={"Authorization": "Bearer sk-eesi-..."},
    timeout=60,
)
response.raise_for_status()
open("preview.mp3", "wb").write(response.content)

To hear arbitrary text in a voice, use text to speech with voice set to the voice id.

Get started

Speech

Realtime

Telephony

Care

Platform

When to use it

Sound like one company

Sound like a person

Speak another language as yourself

Just need a voice?

The invariant

Clone a voice

Built-in voices

Design a voice

Preview

Going further

Voice cloning best practices

Text to speech

Realtime

​When to use it

Sound like one company

Sound like a person

Speak another language as yourself

Just need a voice?

​The invariant

​Clone a voice

​Built-in voices

​Design a voice

​Preview

​Going further

Voice cloning best practices

Text to speech

Realtime

When to use it

The invariant

Clone a voice

Built-in voices

Design a voice

Preview

Going further