Audio

The audio API covers synchronous text-to-speech (TTS), prerecorded speech-to-text (STT), and real-time WebSocket variants. Choose a model from the audio model catalog; each provider supports a different set of voices, languages, formats, and limits.

All HTTP requests use the same bearer-key authentication as the text APIs. TTS returns binary audio, while STT returns JSON or plain text. Audio capability is selected by the model ID; a TTS model cannot be used for transcription and vice versa.

Text-to-speech

POST /v1/audio/speech returns an audio binary stream:

curl -X POST "https://api.llmapi.ai/v1/audio/speech" \
  -H "Authorization: Bearer $LLM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"sonic-3","input":"Welcome to LLM API","voice":"Sarah"}' \
  --output welcome.mp3

The default output is MP3. Use response_format for a format supported by the selected model. The response has an audio content type and should be written to a file or streamed directly to a player; it is not a JSON wrapper.

TTS fields

Field	Description
`model`	TTS model ID, such as `sonic-3` or an ElevenLabs model.
`input`	Text to synthesize. Maximum length depends on the model.
`voice`	Voice ID/name supported by the selected provider.
`response_format`	Output format such as MP3 or WAV, when supported.
`speed`	Playback speed, where supported.
`language`	Language hint, where supported.
`instructions`	Accent or style steering for models that expose it.
`stability`, `similarity_boost`	ElevenLabs voice controls from 0 to 1.

Provider-specific fields are forwarded only when the selected model supports them. Query the audio model guide for voices and limits before exposing these options in a user interface.

Speech-to-text

POST /v1/audio/transcriptions accepts a multipart upload and returns a transcription:

curl -X POST "https://api.llmapi.ai/v1/audio/transcriptions" \
  -H "Authorization: Bearer $LLM_API_KEY" \
  -F "model=ink-whisper" \
  -F "file=@recording.mp3"

The default JSON response is:

{ "text": "Welcome to our platform. We are glad to have you here." }

STT fields

Field	Description
`model`	STT model ID, such as `ink-whisper` or an AssemblyAI model.
`file`	Required audio upload in multipart form data.
`language`	ISO-639-1 language code, when known.
`response_format`	`json` or `text` for all models; `verbose_json` for Whisper models.
`timestamp_granularities`	Set to `word` for Whisper word timestamps with `verbose_json`.
`use_multi_channel`	Transcribe channels independently for models that support it.
`multichannel_output_style`	`separate` or `combined`; combined output requires `verbose_json`.

Use response_format=text when you only need the transcript string. Use JSON or verbose JSON when you need metadata, timestamps, or channel results. The HTTP body limit is 15 MB; accepted codecs, duration, and channel limits remain model-dependent.

Real-time WebSockets

Streaming TTS

Connect to GET /v1/audio/speech/ws, then send JSON messages such as:

{ "type": "tts", "model": "sonic-3", "input": "Hello", "voice": "Sarah" }

The server sends messages such as { "type": "audio", "data": "<base64>" }. Send { "type": "cancel" } to stop the current generation. Every request on one connection must resolve to the same provider.

Streaming STT

Connect to GET /v1/audio/transcriptions/ws and configure the stream first:

{
  "type": "configure",
  "model": "ink-whisper",
  "encoding": "pcm_s16le",
  "sample_rate": 16000
}

After configuration, send raw PCM binary frames. The server emits transcript updates such as:

{
  "type": "transcript",
  "text": "Hello there",
  "is_final": true,
  "words": []
}

Full-duplex voice

Streaming TTS and STT each run in one direction. For a speech-to-speech session where the model listens and answers on the same socket, use Live voice sessions (GET /v1/live/sessions).

Use the Authorization header for either WebSocket endpoint. Streaming TTS also accepts the key as a token query parameter. Streaming sessions include budget protection and can terminate when estimated cost reaches 90% of the remaining balance.

See Audio Streaming for connection examples and message details. WebSocket connections accept the bearer token in the Authorization header; streaming TTS also supports a token query parameter.