Models

Available models

Below is every Speechify text-to-speech model live today. For the authoritative list at runtime, call GET /v1/audio/models — it always returns exactly the models your workspace can use.

Pass the model ID as the model parameter in your API calls. If omitted, the API defaults to simba-3.0; we recommend explicitly setting model: "simba-3.2" on English-only integrations for the lowest TTFB and richest expressivity.

What changed in Simba 3.2, which voices it serves and how to move an existing simba-3.0 integration over is the Simba 3.2 announcement; this page stays the reference for what is live.

Simba 1.6 and your API version

simba-english and simba-multilingual - the Simba 1.6 pair - are being withdrawn in two steps:

If you use either model, nothing breaks on 2026-11-21, but what you hear changes. Moving to a current model yourself before then lets you choose the model and test the voice first: Migrating off Simba 1.6 is the step-by-step guide. Pinning your workspace’s API version to a date before 2026-09-21 keeps the Simba 1.6 training in the meantime, with no code change - see API Versioning for how to read and set it. The pin holds the old training until 2026-11-21 and no further.

New workspaces are pinned to the current version at signup, so a new integration cannot select the pair at all. Use simba-3.2 for English and simba-3.0 for the languages it covers.

Check which languages you actually send before you plan anything. Most workspaces on simba-multilingual synthesize only English, German, Spanish, French, Italian or Portuguese - all covered by simba-3.0, streaming-native and lower latency. If that is you, migrating is a one-line change.

If you do synthesize outside those six languages, broader multilingual coverage is coming on the current model generation before 2026-11-21, and it is the model simba-multilingual is served by from that date. Talk to us if you want to test it earlier.

from speechify import Speechify
client = Speechify(
    token="YOUR_TOKEN_HERE",
)
client.audio.speech(
    audio_format="mp3",
    input="Hello! This is the Speechify text-to-speech API.",
    model="simba-3.2",
    voice_id="geffen_32",
)

Listing models via the API

Fetch the current set of selectable models at runtime instead of hardcoding the table above. GET /v1/audio/models returns, in its models array, every model ID you can pass as the model parameter on the synthesis endpoints. It marks the default (the model used when model is omitted) and the recommended model, and describes each one. Drive a model picker from this response so it stays current as models are added or the recommendation changes.

from speechify import Speechify
client = Speechify(
    token="YOUR_TOKEN_HERE",
)
client.models.list()

{
  "models": [
    {
      "id": "simba-3.0",
      "name": "Simba 3.0",
      "default": true,
      "recommended": false,
      "deprecated": false,
      "description": "Streaming-native model serving both English and multilingual synthesis under one id: English, German, Spanish, French, Italian and Portuguese, routed by the request `language`. The default when a request omits `model`.",
      "languages": [
        "en",
        "de-DE",
        "es-ES",
        "es-MX",
        "fr-FR",
        "it-IT",
        "pt-BR"
      ],
      "endpoints": [
        "/v1/audio/speech",
        "/v1/audio/stream",
        "/v1/audio/stream/with-timestamps"
      ],
      "english_voices_only": false,
      "curated_voices": false
    },
    {
      "id": "simba-3.2",
      "name": "Simba 3.2",
      "default": false,
      "recommended": true,
      "deprecated": false,
      "description": "Streaming-native model with the lowest time-to-first-byte and richest expressivity, English only today. Serves every English voice in the catalog, your workspace's own cloned voices included. Use `simba-3.0` for any other language.",
      "languages": [
        "en"
      ],
      "endpoints": [
        "/v1/audio/speech",
        "/v1/audio/stream",
        "/v1/audio/stream/with-timestamps"
      ],
      "english_voices_only": true,
      "curated_voices": false
    }
  ],
  "dialogue_models": [
    {
      "id": "simba-dialogue-1.0",
      "name": "Simba Dialogue 1.0",
      "default": true,
      "recommended": false,
      "deprecated": false,
      "description": "Multi-speaker model that renders a speaker-attributed script as one conversation with natural turn-taking.",
      "languages": [
        "en"
      ],
      "endpoints": [
        "/v1/audio/dialogue"
      ],
      "english_voices_only": true,
      "curated_voices": false
    }
  ]
}

Each entry carries the model id, a human-readable name and description, a default flag, a recommended flag (the model we suggest for new integrations, which is distinct from the default - the default accepts every voice in every supported language, while the recommended model may be English-only), a deprecated flag, the languages it can synthesize (BCP-47 locale strings matching the language parameter), the endpoints it is valid on, and the curated_voices and english_voices_only flags. deprecated marks a legacy model - a cue to de-emphasise it in a picker and steer new integrations elsewhere. A model being withdrawn carries two more fields: retired_at, the API version at which it stops being selectable (present only while your workspace is pinned below it, because at or after it the model is absent from this response entirely), and sunset_at, the date its own training stops serving the id on every version. Read them together - retired_at is what a pin defers, sunset_at is when that stops working. The list is always exactly what your workspace can call, so a picker driven off it never offers a model your synthesis request would reject. These values reflect current support and can change over time - a model may gain languages, for example - so read them at runtime rather than caching them: because the response shape is stable and clients ignore unknown fields, both changing values and future new fields are picked up without breaking existing integrations.

Voice cloning is supported by both Simba 3 models and requires a paid plan; see Voice Cloning.

Simba 3.2

Streaming-native flagship model with the lowest TTFB (time to first byte) and richest expressivity. Recommended for new English integrations.

• Optimized for real-time streaming with the lowest startup latency
• Richer expressive range than earlier Simba generations
• Full support for SSML and emotion control
• Serves every English voice in the catalog, your workspace’s own clones included
• English only; a non-English voice returns 400. Use simba-3.0 for the other supported languages
• Zero-shot voice cloning is supported: your workspace’s own cloned voices work here like any catalog voice — see Voice Cloning

Simba 3.0

Streaming-native model covering six languages across seven locales: English, German, Spanish (Spain and Mexico), French, Italian, and Brazilian Portuguese. The model a request resolves to when it omits model.

• Officially supports en-*, de-DE, es-ES, es-MX, fr-FR, it-IT and pt-BR. See Language Support
• Set the language parameter to pick the language; when omitted, the voice’s own locale decides
• English and non-English are served by two separate trainings, but that routing is internal: the model ID you pass is always simba-3.0
• Languages outside the supported set often work but are not validated; simba-multilingual covers the full 30+ locale set on a pinned API version, and from 2026-11-21 is served by our current multilingual model
• Prefer simba-3.2 for English-only integrations; it has lower TTFB and richer expressivity
• Full support for SSML and emotion control
• Zero-shot voice cloning works self-serve, in every language the model supports - see Voice Cloning

Simba Multilingual

Supports multiple languages, including mixing languages within a single sentence.

• 31 locales covering 30 distinct languages live today
• Automatic language detection when the language parameter is omitted
• Zero-shot voice cloning works across all supported languages
• Fine-tuned voice cloning available (contact sales)

See Language Support for the full list.

Simba English

Kept for integrations that name it explicitly.

• Full support for SSML and emotion control
• Zero-shot voice cloning from short audio samples
• Fine-tuned voice cloning from hours of speaker audio (contact sales)
• Prefer simba-3.2 for new integrations using the built-in voice catalog; a new workspace cannot select this model at all

Voice cloning

Simba English and Simba Multilingual support two tiers of voice cloning; Simba 3.0 and Simba 3.2 both support zero-shot cloning self-serve:

A cloned voice works on simba-3.0 with no approval step, and on simba-english / simba-multilingual where your API version still offers them; from 2026-11-21 those ids are served by our current models, which take cloned voices too. simba-3.2 cloning is self-serve for every workspace, with no enablement step: an English clone you own works there, and you pass the voice ID exactly as you would a stock voice. (simba-3.2 is English-only, so a non-English clone returns 400 — use simba-3.0 for those.)

See Voice Cloning for implementation details.

FAQ