ElevenLabs Audio

Paths, request bodies, and responses follow ElevenLabs' official Text to Speech and stream-input docs. A model does not necessarily offer every endpoint: check the protocols shown in the model marketplace (last column above). Otherwise HTTP answers 400 model_protocol_not_offered and the WebSocket handshake answers 403.

HTTP speech synthesis

curl "https://api.modelgo.ai/v1/text-to-speech/JBFqnCBsd6RMkjVDRZzb?output_format=mp3_44100_128" \
  -H "Authorization: Bearer $MODELGO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model_id": "<model>",
    "text": "Hello from ModelGo.",
    "voice_settings": { "stability": 0.5, "similarity_boost": 0.75 }
  }' \
  --output speech.mp3

Query parameters such as optimize_streaming_latency and enable_logging are accepted but have no effect.

• The two endpoints without timestamps return audio bytes directly (for example audio/mpeg); /stream sends them as they are generated.
• /with-timestamps returns JSON: audio in audio_base64 (Base64), and alignment / normalized_alignment with characters, character_start_times_seconds, and character_end_times_seconds for every character:

{
  "audio_base64": "<base64 mp3 bytes>",
  "alignment": {
    "characters": ["H", "i"],
    "character_start_times_seconds": [0.0, 0.12],
    "character_end_times_seconds": [0.12, 0.25]
  },
  "normalized_alignment": { "characters": ["H", "i"], "character_start_times_seconds": [0.0, 0.12], "character_end_times_seconds": [0.12, 0.25] }
}

• /stream/with-timestamps returns a run of the JSON objects above (not SSE); parse them one object at a time.
• Billing counts the Unicode characters of text, regardless of output format or audio length; unit prices are in the console's model list.
• Authenticate with Authorization: Bearer <API key>; x-api-key and the xi-api-key header that the official ElevenLabs SDKs send also work, so an official SDK only needs its base URL changed to https://api.modelgo.ai and a ModelGo API key. When several auth headers are present, Authorization wins. Errors use the platform's unified error format; see Errors & Rate Limits.

WebSocket input streaming

For text that arrives in pieces, such as LLM output being spoken as it is generated: the client sends text in chunks and the server returns audio in chunks.

Client messages (JSON text frames):

{"text": " ", "voice_settings": {"stability": 0.5, "similarity_boost": 0.8}}
{"text": "Hello, this is an input streaming test. ", "flush": true}
{"text": ""}

• The first message initializes the session. Its text is usually a single space and it may carry voice_settings (only stability and similarity_boost take effect).
• Each later message appends text; flush: true asks for the text received so far to be synthesized now.
• {"text": ""} ends the input. The server sends the remaining audio, then {"isFinal": true}, and closes with 1000.
• generation_config, try_trigger_generation, pronunciation dictionaries, and similar fields are ignored.

Server messages:

{"audio": "<base64 audio chunk>", "isFinal": null, "normalizedAlignment": {…}, "alignment": {…}}
{"isFinal": true}

audio may be an empty string; skip it.

Handshake errors (before the upgrade, as an HTTP status with a body like {"error":{"type":"invalid_request_error","code":"…","message":"…"}}):

With first-message auth the connection is already upgraded, so these refusals arrive as an error message followed by a close (1011 for the 5xx cases, 1008 otherwise).

Errors during a session: the server first sends {"message": "…", "error": "<code>", "code": <close code>}, then closes with that code:

Billing counts the Unicode characters of the text actually sent for synthesis on the connection (the initializing space, the end marker, and whitespace-only keep-alive messages without flush are free). 5000 characters are reserved when the socket opens and the actual count is settled when it closes; nothing is charged if no character was sent. The platform keeps the provider connection alive while the client is idle, so the client does not need to send keep-alive messages.