Video voiceover
Narration for explainers, ads, social clips and product demos from a script, with a tone that matches the piece and no studio booking. Media teams turn articles and posts into audio at newsroom speed.
SpeechifyAI · Text to speechText-to-speech API.
Text-to-speech API.
Send text, get speech.
Send text to one REST endpoint and get speech back on Simba 3.5, in 48 languages: a whole file with word timings, or a stream.
500K free characters to start. Commercial use, no credit card.
The requestsimba-3.5
bash
curl -X POST https://api.speechify.ai/v1/audio/speech \
-H "Authorization: Bearer $SPEECHIFY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"input": "Ship production voice in a few lines of code.",
"model": "simba-3.5",
"voice_id": "sabrina",
"audio_format": "mp3"
}'What comes back200 OK
json
{
"audio_data": "SUQzBAAAAAAAI1RTU0UAAAAPAAADTGF2ZjYw...",
"audio_format": "mp3",
"billable_characters_count": 45,
"speech_marks": {
"chunks": [
{ "type": "word", "value": "Ship",
"start_time": 0, "end_time": 281 },
...
]
}
}- Simba 3.5: 48 languages under one model id, every catalog voice in each
- POST /v1/audio/speech and POST /v1/audio/stream
- Word timings, SSML and 5 audio formats
- Voice cloning on paid plans
- 500K free characters a month, commercial use
How it works
A voice, a model and your text.
One REST API with an API key. Choose who speaks and which model reads, send the words, and get audio back.
Pick a voice
GET /v1/voices
Every voice comes back with its
id, itslocaleand themodelsit serves, your cloned voices first. For the best results, use a voice native to the language. A voice keeps the accent of the recording it was made from.Pick a model
simba-3.5
The model we recommend: 48 languages under one model id, with every catalog voice in each. Set
languageto your text's language, for exampleen-USorja-JP.Send the text
POST /v1/audio/speech
Returns the whole file and its word timings in one JSON response. For audio as it is generated, send the same request to
POST /v1/audio/stream, up to 20,000 spoken characters a request.
What you getThree models on one API.
Three models on one API.
Simba 3.5 is the one to build on.
- Model
- Simba 3.5Recommended
- Model id
simba-3.5- Languages
- 48 languages
- Voice cloning
- Paid plans
- Per 1M characters, Starter
List price $20launch price $10 · Launch price until 21 Nov
- Model
- Simba 3.2Older model
- Model id
simba-3.2- Languages
- English
- Voice cloning
- Paid plans
- Per 1M characters, Starter
- $10
- Model
- Simba 3.0Older model
- Model id
simba-3.0- Languages
- 6 languages
- Voice cloning
- Paid plans
- Per 1M characters, Starter
- $10
Simba 3.5 lists at $20/$16/$12 per 1M characters on Starter/Pro/Scale, billed at $10/$8/$6 until 21 Nov 2026. Simba 3.0 and 3.2 bill $10/$8/$6 per 1M characters. The Free plan includes 500K characters a month with commercial use.
Streaming or whole file
Stream audio that plays before the sentence has finished generating, or take the complete file in one response.
Word timings
Speech marks give every word's start and end in milliseconds, for captions, highlighting and sync.
Learn moreSSML
Standard tags direct the read: break for pauses, prosody for speaking rate, pitch and volume, sub for pronunciations. The tags are not billed.
48 languages
Simba 3.5 reads all 48 under one model id, and every catalog voice speaks each of them. For the best results, use a voice native to the language. A voice keeps the accent of the recording it was made from.
Learn moreAudio formats
audio_format picks the container: wav, mp3, ogg, aac or pcm, each at 24kHz except wav at 48kHz. output_format sets the rate and bitrate too, such as wav_24000, or ulaw_8000 for telephony. WAV is whole-file only.
Learn more900+ voices
One catalog, each voice listing the models it serves, plus the voices you clone.
Learn moreVoice cloningClone a voice.
Clone a voice.
Use it like any other.
On Starter and every paid plan above it, a short sample and the speaker's verified consent make a voice of your own. Speech in it bills from the same balance, at the rate of the model you call.
Record the speaker
10-30 seconds of speech
A clean sample of the person whose voice you are authorized to clone, in the language you will generate. They also read aloud the phrase
POST /v1/voices/consent-challengesreturns, which verifies their consent.Create the voice
POST /v1/voices
Send the sample, the consent recording and the challenge
id. The response is your new voice, with itsid.Speak with it
simba-3.5
Pass that id as
voice_idtoPOST /v1/audio/speechor/v1/audio/stream, like any catalog voice, in any of the 48 languages.
Listening labVoices you can hear.
Voices you can hear.
Before you integrate.
Prepared Simba 3.5 samples across language and voice identity. Compare the recordings, then take your own text to the playground.
Simba 3.5 · Voice cloning
A new script.
Your voice.
Start with a consented reference recording. Create a reusable voice for the words that come next.
Explore voice cloning01 / Reference
Original speaker
Ready to listen0:00
02 / Clone
Cloned voice
Ready to listen0:00
Two different recordings. Generated sample: Simba 3.5. Only clone a voice you have permission to use.
Simba 3.5 · Multilingual synthesis
Make yourself
understood.
Hear one line in 6 of Simba 3.5's 48 languages, each read by a voice native to it.
Explore language supportEnglish
en-GB · Imogen
German
de-DE · Henrik
Spanish
es-ES · Lola
French
fr-FR · Augustin
Italian
it-IT · Vittoria
Portuguese
pt-BR · Letícia
Prepared Speechify samples. Choose a sample, then press play.Try your own text
Build with these voicesFree tier, no credit card.
The signal
Measured in blind listening.
The Simba 3.5 fact sheet's English results against Simba 3.2, as of October 2026. Fact sheet (PDF)
Latency on Simba 3.2, the English model
In our own open round on 9 Oct 2026, timed from a client in US East over warm connections, Simba 3.2 reached first audible audio in a median of 47 ms over the WebSocket input stream and 64 ms over the HTTP stream. Two independent benchmarks time it from their own runners: Coval measured a 109 ms median time to first audio over the 24 hours to 9 Oct 2026 11:00 UTC, and Voice Arena read 123 ms p50 for its "Simba 3.2 Real-time" row on its US English board on 16 Oct 2026. Their readings include the network to each runner and any silence before the first sound, so they read higher than our production first byte, and neither is its source. The dated snapshots: our round, Coval, Voice Arena.
- win rate against Simba 3.2 in blind side-by-side listening, all content (95% CI 60-67%)
- 63%
- Simba 3.2’s speech throughput per serving instance
- 1.8×
- languages under one model id on Simba 3.5
- 48
- Simba 3.2 on Artificial Analysis's blind-listening leaderboard
- Evaluated
Use casesOne speech layer.
One speech layer.
Many products.
The endpoint stays the same. The voice, timing, format, and delivery adapt to the experience around it.
Podcasts
A written script becomes a produced episode: assign a voice per speaker, synthesize the lines, assemble the audio. Solo shows, co-host banter and interview formats without coordinating a recording.
Audiobooks and publishing
A manuscript becomes a finished audiobook with one consistent voice across every chapter, and a publisher's whole catalog runs through the same pipeline instead of a studio session per title.
Accessibility and read-aloud
Articles, documents, app screens and web pages spoken aloud with a follow-along cursor from speech marks, for readers with dyslexia, low vision or a preference for audio.
E-learning and training
One consistent narrator for every module, captions from speech marks, and multilingual synthesis with Simba 3.5 across 48 languages. A voice keeps the accent of the recording it was made from, so pick a voice native to each language. A revision is a text edit and a re-render, not a re-recording.
Games
Thousands of lines and dozens of characters voiced from a dialogue script without a casting session, and lines generated at runtime for scenes that are not fixed at ship time.
IVR and phone prompts
Menu options, hold messages, announcements and after-hours notices generated from text in the format a phone system expects, updated in seconds rather than studio sessions.
Spoken notifications
A delivery update, an appointment reminder, a system warning or an in-app cue becomes a clear spoken message, synthesized from whatever the alert says rather than a fixed library of recordings.
FAQ
Frequently asked questions
What is a text-to-speech API?
A text-to-speech API turns written text into spoken audio over a network request. You send text, a voice and a model; it returns audio, either as a complete file or as a stream. On SpeechifyAI you build on Simba 3.5, which speaks 48 languages under one model id, with SSML control, voice cloning and word-level speech marks on the same endpoints.
How fast is it?
Every Simba model is streaming-native, so audio starts playing before the full clip is generated. Simba 3.2, the English model, returns its first byte in 56 ms at the median and 102 ms at the 90th percentile, measured on the production US East streaming path on 15 Sep 2026; the site rounds that to <100ms. In our own open round on 9 Oct 2026, a client in US East reached first audible audio on it in a median of 47 ms over the WebSocket input stream and 64 ms over the HTTP stream.
Has anyone measured it independently?
Yes. Two independent benchmarks time Simba 3.2 from their own runners to the first audible sound: Coval measured a 109 ms median time to first audio over the 24 hours to 9 Oct 2026 11:00 UTC, and Voice Arena read 123 ms p50 for its "Simba 3.2 Real-time" row on its US English board on 16 Oct 2026. Their readings include the network to each runner and any leading silence, so they sit above our own first byte, which they do not source.
Can it control pacing and pauses?
Yes. Standard SSML directs the timing word by word: break for pauses, prosody rate for speaking rate, and sub for pronunciations, at no extra cost. Simba 3.5, Simba 3.2 and Simba 3.0 also apply prosody pitch and volume, and accept emphasis and the speechify:style emotion tag but do not apply them.
What does it cost?
Paid plans have a monthly subscription and an included USD usage balance. Simba 3.5 lists at $20/$16/$12 per 1M characters on Starter/Pro/Scale, billed at $10/$8/$6 until 21 Nov 2026. Simba 3.0 and 3.2 bill $10/$8/$6 per 1M characters. The Free plan includes 500K characters a month with commercial use. Voice cloning requires a paid plan. See Pricing for subscriptions, included balances, and additional usage.
Which models and languages can I use?
Simba 3.5 speaks 48 languages under one model id, simba-3.5, and every voice in the catalog speaks all 48. The older Simba 3.2 (English) and Simba 3.0 (6 languages across 7 locales) remain available. All three are streaming-native and support voice cloning on paid plans.
Which audio formats are supported?
Five: wav, mp3, ogg, aac, and pcm. The batch endpoint returns a full file including WAV; the streaming endpoint returns chunked mp3, ogg, aac, pcm, or u-law for telephony.
Enterprise
Enterprise-ready when you are.
Start on the free tier today and bring procurement in later. SOC 2 Type II, SSO, a signed DPA and volume pricing are all here when you need them, and teams with heavy traffic or a regulated surface get a Forward Deployed Engineer building alongside them.
Start building in minutes.
Issue a key, paste the request, and hear your first audio come back. Free while you build, and priced per character the day you ship.
500K free characters to start. Commercial use, no credit card. SOC 2 Type II.