Designing voices

Voice design creates a voice from a plain-language description. Send "voice": "design" and an instructions description on text-to-speech — POST /v1/audio/speech (MP3 by default) or POST /tts (WAV by default) on gpu-tts.ollalink.com or gpu-voice.ollalink.com, with your GPU API key in X-NH-GPU-Key. Designed voices speak all 14 languages: English, Hindi, Spanish, French, Chinese, German, Arabic, Brazilian Portuguese, Russian, Japanese, Korean, Italian, Dutch and Vietnamese.

curl https://gpu-tts.ollalink.com/v1/audio/speech \
  -H "X-NH-GPU-Key: $YOUR_GPU_KEY" -H "Content-Type: application/json" \
  -d '{"voice": "design", "language": "en", "response_format": "wav",
       "instructions": "A young woman with a clear Indian English accent, bright and friendly, a tech YouTuber explaining networking to beginners",
       "input": "Today we will configure OSPF on two routers and check the neighbour table."}' \
  --output designed.wav

Request fields

Field	Meaning
voice	design
instructions	Required, up to 400 characters: the description of the voice (see Writing a description)
input	Required, up to 1,000 characters per request. Write it in the language you want spoken; split longer text at sentence ends
language	Optional: en, hi, es, fr, zh, de, ar, pt, ru, ja, ko, it, nl, vi
response_format	mp3 (default on /v1/audio/speech), wav, pcm, opus, aac, flac

The response body is the audio, at 24 kHz like every voice. Headers: X-TTS-Engine: design, X-TTS-Voice: design and X-Audio-Seconds. A short clip typically comes back in a few seconds.

Each request invents a new speaker. The same description gives a similar voice, but not an identical one. A designed voice cannot be saved or reused, so for one consistent voice across many requests use a ready-made voice.

Writing a description

Write one sentence of about 30–45 words in English (Chinese is also accepted for Chinese voices) that covers:

• accent, age and gender;
• pitch and texture of the voice;
• energy and pace;
• the role or scenario.

Describe the sound, not celebrities or brands. For a strong regional accent, name it precisely and add a city or role (“a Bengaluru tech YouTuber”). If the description does not mention a recording setting, the service adds “recorded in a quiet professional studio, clean audio with no background noise and no echo”. In Hindi and Chinese text, numbers read best written as words.

Examples:

• “A young woman with a clear Indian English accent, bright and friendly, a tech YouTuber explaining networking to beginners”
• “A calm middle-aged man with a warm, low voice and a neutral American accent, speaking slowly and clearly like a training narrator”
• “A cheerful young Spanish woman from Madrid with a light, clear voice and a brisk pace, presenting a product demo”

Errors

Status	When
400	no instructions; instructions over 400 characters; input over 1,000 characters; a language that is not offered
503	the voice engine is busy or briefly unavailable — retry shortly

Not offered

• Cloning from your own audio is paused: ref_audio / ref_text answer 400 (Voice cloning).
• Saving a designed voice for later requests is not offered.
• The voice studio routes on gpu-voice (/voice_design, /custom_voice, /clone, /speakers) answer 404 (Voice studio); use "voice": "design" on /v1/audio/speech instead.
• instructions on a ready-made voice answers 400: those voices have a fixed style.