Kokoro Text to Speech — 54 voices, 9 languages

Turns text into natural speech with Kokoro-82M, an open-weight text-to-speech model that punches far above its 82 million parameters. 54 voices across nine languages, delivered as MP3, WAV, Opus, FLAC or AAC.

Paste a list of texts. Get back one audio file per entry, plus a row describing each one.

Input

Field	Type	Default	What it does
texts	array	—	One entry per audio file. A long entry is split at sentence boundaries and joined back into a single file, so a whole article can go in one line. Up to 500,000 characters per run.
voice	select	af_heart	One of 54 voices. The first letter is the accent (a American, b British, e Spanish, f French, h Hindi, i Italian, j Japanese, p Portuguese, z Mandarin); the second is the speaker's gender.
language	select	en-us	Pronunciation rules the text is read with: en-us, en-gb, es, fr-fr, hi, it, ja, pt-br, cmn.
speed	select	1.0	0.5x to 2.0x. 1.0 is the model's natural pace.
audioFormat	select	mp3	mp3, wav, opus, flac, aac.

Match the language to the voice. A Japanese voice reading English under American pronunciation rules sounds wrong in a way no amount of voice-shopping fixes.

Output

The audio goes to the run's key-value store — a dataset record is JSON and cannot hold bytes — and each dataset row carries its URL:

index             position of the text in your input list
audio_url         the generated file
duration_seconds  length of the audio
character_count   characters actually spoken
billed_units      what this row was charged: characters rounded up to the next 1,000
voice             the voice used
language          the pronunciation rules used
speed             the rate it was spoken at
audio_format      the container
sample_rate       24,000 Hz
chunks            how many synthesis chunks the text was split into
text              the text as it was spoken, after splitting and rejoining

Billing

Pay per event, priced per 1,000 characters and rounded up, plus a small actor-start fee covering the cold start and model load. The unit is characters rather than seconds of audio so the price of a job is known before it runs: a 40,000-character article is 40 units whatever the voice does with it.

• An empty entry is not synthesised and not charged.
• A run stops as soon as your spending limit is reached, rather than continuing to generate audio you are not paying for.

Quality notes

• Long text is split at sentence boundaries, never mid-sentence. A cut inside a sentence is audible — the prosody resets and the join clicks.
• The model runs on ONNX Runtime at full precision. The quantised build is roughly a third of the size but noticeably noisier on sibilants.
• Kokoro is expressive but not controllable: there are no SSML tags, no per-word emphasis and no emotion parameter. What you shape it with is punctuation and the voice choice.
• Numbers, abbreviations and units are read as written. Dr. and 3.5 kg are handled; a bare 2026-08-30 is read as digits, so spell out anything you want spoken a particular way.

What it does not do

• No voice cloning. The 54 voices are the voices.
• No streaming. This is a batch actor: it returns files, not a live audio stream.
• No word-level timestamps.