Appearance
Models and voices
Available models
| Model | Type | Endpoints | Notes |
|---|---|---|---|
cadence-tts | Text to speech | POST /v1/audio/speech | Low-latency streaming. wav, mp3 and pcm output |
cadence-stt | Speech to text | POST /v1/audio/transcriptions, WS /v1/realtime | Files and live audio. Multilingual |
Cadence is the xVoice model family. Prices are on xvoice.dev.
This table is a summary. What each model accepts is published on the model itself, as described below, and the API enforces exactly that.
List models
GET /v1/models
python
for model in client.models.list():
print(model.id, model.type, model.capabilities)js
for await (const model of client.models.list()) {
console.log(model.id, model.type, model.capabilities)
}bash
curl https://api.xvoice.dev/v1/models \
-H "Authorization: Bearer $XVOICE_API_KEY"Add ?type=stt or ?type=tts to filter. GET /v1/models/{model_id} returns one model.
The model object
Models are standard OpenAI model objects, so the SDKs parse them as usual, with xVoice's fields added:
json
{
"id": "cadence-stt",
"object": "model",
"created": 1758672000,
"owned_by": "xvoice",
"name": "Cadence STT",
"type": "stt",
"description": "Multilingual speech-to-text for recorded audio.",
"capabilities": ["batch", "realtime"],
"supported_languages": ["ar", "de", "en", "es", "fr", "..."],
"status": "available",
"public_version": "2026-09-17",
"speech_options": null
}| Field | Description |
|---|---|
type | stt or tts. Decides which endpoints accept the model |
capabilities | Which ways of calling the model are available. See below |
supported_languages | ISO 639-1 codes the model supports |
status | available or deprecated |
public_version | The version requests resolve to today. See Versions |
speech_options | For TTS models, the parameters the model accepts. null for STT models |
capabilities
| Capability | Meaning |
|---|---|
batch | Accepts a complete request: a file to POST /v1/audio/transcriptions, or a non-streamed POST /v1/audio/speech |
streaming | Can stream its output over HTTP |
realtime | Can hold a live session on WS /v1/realtime |
Filter on capabilities instead of guessing from a model's name. It is enforced: opening a realtime session with a model that lacks realtime fails with realtime_not_supported.
speech_options
TTS models publish the speech parameters they accept:
json
{
"response_formats": ["wav", "pcm", "mp3"],
"streamable_formats": ["wav", "pcm", "mp3"],
"speed": {"min": 1.0, "max": 1.0},
"instructions": false
}| Field | Meaning |
|---|---|
response_formats | Formats the model can return |
streamable_formats | Formats the model can stream. A model with any of these has the streaming capability |
speed | The allowed range for speed |
instructions | Whether the model acts on instructions. When false, the field is accepted and ignored |
A request outside these values fails with 400 unsupported_response_format or 400 unsupported_speed, and the message lists what the model accepts. Build format and speed pickers from this object, so a new model works without a code change.
status and deprecation
A deprecated model still serves requests, but will be retired. Move to its replacement. Retired models are removed from the list, and requests naming them fail with invalid_model.
A listed model that temporarily cannot serve requests fails with 404 model_unavailable. That is a problem on our side: retry later, and contact support with the request id if it continues.
Versions
public_version identifies the model version your requests use today. You do not send it, and pinning a version is not available yet. When a model's output changes, the version changes too, so you can quote it when you report a difference.
Voices
GET /v1/voices
Voices for cadence-tts:
| Voice | Gender | Accent | Speaks |
|---|---|---|---|
clara | Female | en-US | en, es |
maya | Female | en-US | en, es |
ethan | Male | en-US | en, es |
noah | Male | en-US | en, es |
oliver | Male | en-US | en, es |
Pass the id as voice in a text to speech request, and pass language when the text is not in the voice's own language. See Languages.
List voices
The OpenAI SDKs have no voices method, so call the endpoint directly:
bash
curl "https://api.xvoice.dev/v1/voices?model=cadence-tts" \
-H "Authorization: Bearer $XVOICE_API_KEY"python
voices = client.get(
"/voices",
options={"params": {"model": "cadence-tts"}},
cast_to=object,
)["data"]js
const { data: voices } = await client.get('/voices', { query: { model: 'cadence-tts' } })json
{
"object": "list",
"data": [
{
"id": "clara",
"name": "Clara",
"locale": "en-US",
"languages": ["en", "es"],
"gender": "female",
"preview_url": null,
"status": "available",
"metadata": {}
}
]
}| Field | Description |
|---|---|
locale | The accent the voice was recorded in, such as en-US |
languages | ISO 639-1 codes the voice can speak. With ?model=, on that model; otherwise on any model |
| Parameter | Description |
|---|---|
model | Only voices that model can use. Use this to build a voice picker |
language | Only voices that can speak this language, such as es |
locale | Only voices with this accent, such as en-US |
A voice works only with the models it is available on. Using it with another model fails with voice_not_supported, so filter by model rather than assuming a voice works everywhere.