Kokoro 82M

The cheapest speech model on the platform by a wide margin — 54 voices across 8 languages.

POST/v1/media/queue

$0.0014/1K chars

Overview

Kokoro 82M is a lightweight, open-weight text-to-speech model from hexgrad. At 82M parameters it is by far the cheapest speech model on YepAPI — roughly 24× cheaper per character than the mid-tier models — while still covering 8 languages with 54 named voices. Voice IDs are prefixed by language and gender (af_ American female, bm_ British male, jf_ Japanese female, and so on).

Property	Value
Model ID	`hexgrad/kokoro-82m`
Alias	`kokoro`
Upstream Model	`hexgrad/kokoro-82m`
Category	Text to Speech
Languages	American and British English, Spanish, French, Hindi, Italian, Japanese, Portuguese, and Chinese
Output	`mp3` (default) or `pcm`
Pricing	$0.0014 per 1,000 characters

Usage

All media models use the async job queue. Submit a job, then poll for the result.

Step 1: Submit Job

const res = await fetch('https://api.yepapi.com/v1/media/queue', {
  method: 'POST',
  headers: {
    'x-api-key': 'YOUR_API_KEY',
    'Content-Type': 'application/json',
  },
  body: JSON.stringify({
    model: 'hexgrad/kokoro-82m',
    prompt: 'Your text to speak goes here.',
    options: { voice: 'af_bella' },
  }),
});
const { data } = await res.json();
// data.jobId — use this to poll for results

curl -X POST https://api.yepapi.com/v1/media/queue \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "hexgrad/kokoro-82m", "prompt": "Your text to speak goes here."}'

Step 2: Poll for Result

const status = await fetch(`https://api.yepapi.com/v1/media/status/${data.jobId}`, {
  headers: { 'x-api-key': 'YOUR_API_KEY' },
});
const { data: job } = await status.json();
// job.status — "pending" | "processing" | "completed" | "failed"
// job.result.audio — { mimeType, base64 } when completed

curl https://api.yepapi.com/v1/media/status/JOB_ID \
  -H "x-api-key: YOUR_API_KEY"

Write the audio to a file:

import { writeFileSync } from 'node:fs';

writeFileSync('speech.mp3', Buffer.from(job.result.audio.base64, 'base64'));

Request Body

Parameter	Type	Required	Description	Default
`model`	`string`	Yes	`hexgrad/kokoro-82m`	—
`prompt`	`string`	Yes	The text to speak (max 50,000 bytes)	—
`options.voice`	`string`	No	Voice identifier	`af_bella`
`options.outputFormat`	`string`	No	`mp3` (default) or `pcm`	`mp3`
`options.speed`	`number`	No	Playback speed multiplier. Honoured only by models that support it	`1.0`

Voices

af_bella is used when options.voice is omitted. 54 voices available (11 of 54 shown):

af_bella
af_heart
af_nova
am_adam
am_michael
bf_emma
bm_george
ef_dora
ff_siwis
jf_alpha
zf_xiaoxiao

Features

Cheapest speech model on the platform by a wide margin
54 voices across 8 languages
Language/gender-prefixed voice IDs (af_, am_, bf_, bm_, ef_, ff_, hf_, if_, jf_, pf_, zf_…)
Open weights

Billing

Speech is billed on the length of the input text, in UTF-8 bytes — for English text one byte is one character. The cost is known before synthesis starts, so the balance check at submit time quotes the exact final charge. Jobs have a $0.01 minimum.

Under the Hood

Kokoro 82M

On this page