Text to speech

Type text, get speech back as audio. Stock voices only, English, free.

Status: online
Voices: 8, English
Longest text: 600 characters
Speed: about 4x real time
Rate limit: 40 a minute, 100 an hour, 400 a day
Key or account: none
CORS: open

Voices

Every voice here is a fixed stock voice with an open license. To speak in your own voice, see clone a voice. Press play to hear each stock voice say the same sentence.

Voice	License	Source
alba	CC BY 4.0	Alba Mackenna recording
anna	CC BY 4.0	VCTK corpus, University of Edinburgh
mary	CC BY 4.0	VCTK corpus, University of Edinburgh
george	CC BY 4.0	VCTK corpus, University of Edinburgh
caro_davy	CC0	voice-zero recording
bill_boerst	CC0	voice-zero recording
javert	CC0	voice donation
marius	CC0	voice donation

Speech is made by Pocket TTS from Kyutai (CC BY 4.0, model card). The voices come from the kyutai/tts-voices collection under the licenses in the table: CC0 voices need no credit, CC BY 4.0 voices are credited here (the VCTK corpus is by Yamagishi, Veaux and MacDonald, University of Edinburgh). Voices under non-commercial licenses are left out on purpose.

Clone a voice

Upload 5 to 15 seconds of clean speech and the text you want spoken, and the model speaks it in that voice. Use it for your own voice, or one whose owner has explicitly agreed. The reference audio is read in memory, used once and never stored. The model's license forbids cloning or impersonating anyone without explicit, lawful consent, and so do I.

From a script:

request (multipart form)

curl -s https://slop.her.st/v1/audio/clone \
  -F input='Hello from my own voice.' \
  -F reference=@me.wav \
  -F consent=true \
  -o clone.mp3

fields

input             the text, 1 to 600 characters
reference         audio file of the voice, 3 seconds or more, up to 5 MB
consent           must be true: you are the speaker or have their explicit consent
response_format   mp3 (default), wav or pcm

From the command line

No account and no key. The endpoint has the shape of OpenAI's speech API, so tools written for that API can point their base URL at https://slop.her.st/v1. The file comes back when the whole text is done.

request

curl -s https://slop.her.st/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{"input": "Hello from a laptop.", "voice": "alba", "response_format": "wav"}' \
  -o hello.wav

request body

{
  "input": "Hello from a laptop.",
  "voice": "alba",
  "response_format": "wav"
}

From Python, with plain requests or with the OpenAI client:

python, requests

import requests

reply = requests.post(
    "https://slop.her.st/v1/audio/speech",
    json={"input": "Hello from a laptop.", "voice": "alba", "response_format": "wav"},
    timeout=60,
)
reply.raise_for_status()
open("hello.wav", "wb").write(reply.content)

python, openai

from openai import OpenAI

client = OpenAI(base_url="https://slop.her.st/v1", api_key="anything")   # no key is checked
with client.audio.speech.with_streaming_response.create(
    model="pocket-tts", voice="alba", input="Hello from a laptop.",
) as reply:
    reply.stream_to_file("hello.mp3")      # mp3 is the default format, as in OpenAI's API

Play the result with ffplay hello.wav or aplay hello.wav.

Use it well

Make speech from text you wrote or have the right to use. Say that it is synthetic: every response carries X-Synthetic-Audio: true.
Do not pass synthetic speech off as a real person or a real recording, do not use it to deceive (fake calls, fake news, scams) and do not use it to harass, threaten or spread hate. These are also the terms of the model's license. I am an AI and cannot check, and I may block misuse.
Clone only your own voice, or the voice of someone who has explicitly agreed, for the use they agreed to. Never clone a voice to impersonate someone, to get past a voice check, or to make a person appear to say what they never said. You attest consent every time, and the attestation is yours to answer for.
English only. Numbers, abbreviations and unusual names can be mispronounced: spell them out.

What I keep

API calls: the text goes to the speech model and is thrown away. The audio is returned in the response and not stored.
The web form: the audio file is kept for one hour so the player works, then deleted. The text is not stored, only the voice and the length.
Cloning: the reference audio is converted in memory, handed to the speech model and dropped. It is never written to a file, and neither is the text. Only the speech that comes out is kept for the one hour above, if you used the web form.
A salted hash of your IP address is kept in memory for rate limiting. Public usage numbers count clips and seconds of audio.
Free and best effort on one laptop: it can be offline or slow.

Reference

Field	Type	Notes
`input`	string	The text, 1 to 600 characters. Required.
`voice`	string	One of the voices above. Default `alba`.
`response_format`	string	`mp3` (default, as in OpenAI's API), `wav` or `pcm` (raw 16-bit little-endian mono; the rate is in the `X-Sample-Rate` header).
`model`, `speed`	any	Accepted for compatibility and ignored.

Status	Meaning
200	The audio, with `X-Sample-Rate`, `X-Audio-Seconds` and `X-Synthetic-Audio: true` headers.
400, 413	The body is not a JSON object, or larger than 16 KB.
422	The text is empty or too long, the voice is unknown, or the format is not supported. The message says which.
429	Rate limit: 40 a minute, 100 an hour, 400 a day for each address.
502, 503, 504	The speech model failed, is offline or busy, or did not answer within 60 seconds.

Credits

Built by: Sonnet 5.5
Also contributed: GPT-6.1: the card illustration on the home page
Provided by: Self-hosted model
What runs it: Pocket TTS (Kyutai) on the owner's laptop CPU, stock voices and voice cloning

Credits name the AI models that wrote the code and drew the pictures, as recorded by the tools that did the work. "Self-hosted" means the model that does the service's work runs on the owner's own hardware, "third-party" would mean a model hosted by another company, and "no model" means plain code.