Simba 3.5: one text-to-speech model for 48 languages

Introducing Simba 3.5. From 21 October 2026, one model id, simba-3.5, speaks 48 languages on every plan, up from the 6 that Simba 3.0 covers, through the endpoints and voices you already use.

It is our recommended model from today. simba-3.0 stays the default when a request omits model, and simba-3.0 and simba-3.2 stay available.

One request

Set model to simba-3.5 and language to the language of your text. This streams German from Amelie, a German voice in the catalog, into an MP3:

curl https://api.speechify.ai/v1/audio/stream \
  -H "Authorization: Bearer $SPEECHIFY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: audio/mpeg" \
  -d '{
    "input": "Ein Modell, achtundvierzig Sprachen.",
    "voice_id": "amelie",
    "model": "simba-3.5",
    "language": "de-DE"
  }' \
  --output speech.mp3

The same model value works on POST /v1/audio/speech, POST /v1/audio/stream/with-timestamps for word timings, and the WebSocket at GET /v1/audio/stream/ws. Multi-speaker scripts on POST /v1/audio/dialogue stay on simba-dialogue-1.0.

The 48 languages

Arabic, Belarusian, Bengali, Bosnian, Burmese, Catalan, Chinese, Croatian, Czech, Danish, Dutch, English, Finnish, French, Georgian, German, Greek, Gujarati, Hebrew, Hindi, Hungarian, Indonesian, Italian, Japanese, Kazakh, Korean, Latin, Malay, Marathi, Norwegian, Persian, Polish, Portuguese, Romanian, Russian, Serbian, Slovak, Slovenian, Spanish, Swedish, Tagalog, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu and Vietnamese.

Any regional variant of each is in scope, so pt-BR and pt-PT are both Portuguese. Any catalog voice can read any of them, and so can a voice you cloned on a paid plan. A voice keeps the accent of the recording it was made from, so for a native accent pick a voice from the language itself. The language table in the docs gives the code to send for each.

Measured

English results against Simba 3.2, as of October 2026: blind side-by-side listening by US and UK English listeners, and speed under production load. In blind listening tests in the six locales both models speak, native listeners prefer Simba 3.5 to Simba 3.0 Multilingual in 62% of comparisons, and each Simba 3.5 serving instance delivers on average 1.4× the speech throughput of a Simba 3.0 Multilingual instance. The figures and their notes are in the Simba 3.5 fact sheet (PDF).

What it costs

Simba 3.5 lists at $20 per 1M characters on Starter, $16 on Pro and $12 on Scale, twice the rates of Simba 3.0 and Simba 3.2. Until 21 November 2026 it bills at their rates instead: $10, $8 and $6.

A plan’s month is one balance in dollars, drawn at the rate of the model you call. At list price it buys 950K characters of Simba 3.5 on Starter and 250K on Free. Until 21 November the launch price buys twice that: 1.9M and 500K.

Every model bills the spoken characters of a request, with SSML tags not counted, so a request costs the same number of characters on simba-3.5 as on simba-3.2.

Switching from Simba 3.2

1. Change model from simba-3.2 to simba-3.5. Voice IDs, output formats, speech marks and the SSML that Simba 3.2 applies, <break>, <prosody rate> and <sub alias>, carry over, and no API version change is needed.
2. Set language on every request whose text is not in the voice’s own language. Without it, the voice’s locale decides.
3. Keep simba-3.2 on any call you have not moved yet: both ids answer on the same endpoints, so you can move one call at a time.

Every model side by side is on /models. The reference is Models in the docs, and the release note is the 21 October 2026 changelog entry.