Models

KittenASR Enhanced listens; Kitten TTS 2 and three KittenTTS 0.8 models speak. Model IDs are case-sensitive, so send the exact ID shown below.

Speech to text

Model ID	Input	Output
kittenasr-enhanced-preview	Speech audio	Clean text plus emotion-aware transcript metadata

The same model ID is used for upload and realtime STT.

GET /v1/models reports this model with the identifier kitten-asr-pro-enhanced. That identifier and kittenasr-enhanced-preview name the same hosted model and are interchangeable in every request.

For migrations from OpenAI clients, whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-realtime-whisper are accepted compatibility aliases for the same hosted KittenASR Enhanced model. New integrations should use kittenasr-enhanced-preview; the aliases are not separate models.

Text to speech

Two model families speak. Kitten TTS 2 is the current hosted model: 47 voices, nine languages, voice cloning, emotion markup, and decoding controls. KittenTTS 0.8 is the earlier, much smaller family: three sizes, eight English voices, and none of the Kitten TTS 2 features. Both use the same endpoints and the same request shape, so switching is a change of model and voice.

Choose a model

If you need	Use
The best quality, more voices, or anything other than English	kitten-tts-2-latest
Voice cloning, emotion and sound tags, or decoding controls	kitten-tts-2-latest
The smallest model, or the same voices you run on device with the KittenTTS SDKs	A KittenTTS 0.8 model

A request without model is served by kitten-tts-mini-0.8, not Kitten TTS 2, so always name the model you want.

Model ID	Family	Size	Built-in voices	Languages
kitten-tts-2-latest	Kitten TTS 2	1.7B ternary	47 voices; default eleanor_somber_female_32	English, Arabic, German, Spanish, French, Italian, Portuguese, Russian, Simplified Chinese, Hindi
kitten-tts-mini-0.8	KittenTTS 0.8	80M	8 classic voices; default Bella	English
kitten-tts-micro-0.8	KittenTTS 0.8	40M	8 classic voices; default Bella	English
kitten-tts-nano-0.8	KittenTTS 0.8	15M	8 classic voices; default Bella	English

kitten-tts-2-latest always serves the current Kitten TTS 2 release, so its speed and output can improve between releases.

Feature support

Feature	Kitten TTS 2	KittenTTS 0.8
Complete response, binary and SSE streaming	Yes	Yes
Realtime WebSocket input streaming	Yes	Yes
Output formats mp3, wav, flac, aac, opus, pcm	Yes	Yes
speed from 0.25 to 4.0	Yes	Yes
Up to 40,000 input characters per HTTP request	Yes	Yes
Numbers, dates, and units read as words	English voices, and English text on the language voices	English
Voice cloning: one-request reference, saved voices, shared voices	Yes	No
List voices with GET /v1/voices	Yes	No; the eight voices are fixed
Emotion and sound tags, (((emphasis)))	Yes	No; bracketed text is read aloud
Decoding controls: mode, temperature, top_p, top_k, min_p, max_new_tokens	Yes	No; the fields are ignored

Voice names belong to one family: a Kitten TTS 2 voice is rejected by a 0.8 model and a 0.8 voice by Kitten TTS 2. The eight 0.8 speakers are also available in Kitten TTS 2 as bella_10 through leo_17.

For pricing, see the KittenML platform.

kitten-tts-0.9 still works

Kitten TTS 2 was previously published as KittenTTS 0.9, with the model ID kitten-tts-0.9. That ID is now a legacy alias of kitten-tts-2-latest: it is still accepted everywhere a model is named (request bodies, form fields, and the model query parameter), and it selects the same model with the same voices and audio. GET /v1/models lists both IDs. A response names the model with the ID you sent (for example, in the X-Model header), so compare model names against both IDs, never just one. New integrations should use kitten-tts-2-latest.

List models

curl --fail-with-body -sS \
  https://api.kittenml.com/v1/models \
  -H "Authorization: Bearer $KITTENML_API_KEY"

Use the catalog above as the supported inference contract. The model-list endpoint currently requires a key with ASR access and paid-ASR eligibility.