Audio / Speech API

There are two ways to do text-to-speech: the synchronous, OpenAI-compatible POST /v1/audio/speech (Turkish ema/ema-lightning, free beta; below) and asynchronous generation with ElevenLabs TTS models (POST /v1/audio/generate; 29 languages including Turkish). For speech-to-text: use the POST /v1/audio/transcriptions endpoint with a :tr speech recognition model (multipart form), or send openai/gpt-audio or google/gemini-2.5-flash to /v1/chat/completions with input_audio. Audio output (synchronous): on the chat endpoint send modalities: ["text","audio"] + audio: {"voice":"alloy","format":"pcm16"} with stream: true; supported by openai/gpt-audio and openai/gpt-audio-mini. Audio arrives base64-encoded in the SSE stream under delta.audio.data. Note: in this mode format must be pcm16 and stream: true is required.

Synchronous Text-to-Speech: POST /v1/audio/speech (Free Beta)

Parameters

Output Formats

Examples

curl https://api.onysoft.com/v1/audio/speech \
  -H "Authorization: Bearer sk-ony-YOUR-KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ema/ema-lightning",
    "input": "Merhaba, siparişiniz yola çıktı ve yarın teslim edilecek.",
    "voice": "ema",
    "response_format": "mp3"
  }' \
  --output konusma.mp3

from openai import OpenAI

client = OpenAI(base_url="https://api.onysoft.com/v1", api_key="sk-ony-YOUR-KEY")

with client.audio.speech.with_streaming_response.create(
    model="ema/ema-lightning",
    voice="ema",
    input="Randevunuz 15 Ekim saat 14:30 için onaylandı.",
    response_format="wav",
) as yanit:
    yanit.stream_to_file("randevu.wav")

Quotas and Limits

• Free beta quota (default) per identity: 100 requests, 100,000 characters and 120 minutes of audio per day. Identity: the account for user keys (all keys of the account share it), the key itself for project and partner keys. All customer and project keys of one partner can produce at most 6 hours of audio per day in total. The free beta also has a limited server-wide daily capacity. Quotas reset at 00:00 Turkey time.
• Successful responses carry the remaining quota in headers: X-Quota-Requests-Remaining, X-Quota-Characters-Remaining, X-Quota-Audio-Seconds-Remaining and X-Quota-Reset (Unix time). These headers, Retry-After and X-Request-Id are also readable from browser JavaScript via Access-Control-Expose-Headers.
• When a quota is used up you get 429 tts_daily_limit with a Retry-After header holding the seconds until midnight and X-Should-Retry: false (SDKs do not retry immediately). Counted against the quota: successful requests, and requests that timed out (504 tts_timeout) or failed (500 tts_generation_failed) after synthesis started. Validation errors, 429 rate_limited responses and 503 responses while the service is down are not counted.
• At most 2 requests per identity (the account for user keys, the key for project and partner keys) are processed at the same time; extra requests get 429 rate_limited. Contact support for a higher quota or concurrency.
• Read-aloud limit: the text is measured with numbers and symbols spelled out. A request whose estimated audio exceeds 20 minutes gets 400 input_too_long without waiting in the queue; the spoken_characters, estimated_audio_seconds and max_audio_seconds fields show how much to split it. Such requests do not count against the quota.
• limit_type values: tts_daily_requests, tts_daily_characters, tts_daily_audio_seconds, tts_partner_daily_audio_seconds, tts_global_daily_audio_seconds (429 tts_daily_limit); tts_concurrency, tts_busy (429 rate_limited).

Error Codes

Audio Generation (TTS)

Quick Example (Turkish TTS, Rachel)

curl -X POST https://api.onysoft.com/v1/voice/generate \
  -H "Authorization: Bearer sk-ony-YOUR-KEY" \
  -d '{
    "model": "elevenlabs/text-to-speech-multilingual-v2",
    "prompt": "Merhaba, ben Rachel. Onysoft AI üzerinden konuşuyorum.",
    "voice": "Rachel"
  }'

Supported Voices (case-sensitive)

Female: Rachel (default), Alice, Aria, Charlotte, Jessica, Laura, Lily, Matilda, Sarah
Male: Brian, Bill, Callum, Charlie, Chris, Daniel, Eric, George, Liam, Roger, Will

Audio Models

* ElevenLabs models may have temporary errors on the generation service side; failed generations are not charged. For Turkish text-to-speech you can also use the synchronous POST /v1/audio/speech endpoint.

Status Lookup

curl https://api.onysoft.com/v1/audio/status/vtask_xyz789 \
  -H "Authorization: Bearer sk-ony-YOUR-KEY"

For all parameters (stability, similarity_boost, style) and TTS documentation, see video-api → Audio Models.