Alania-2 · Text to speech

Our new Turkish voice: Alania-2.

We trained Alania-2 from scratch on our own Turkish data. It speaks at 48 kHz and comes with 11 built-in voices. You can tell it how to speak in a sentence, or describe a new voice in words. It runs on the same API and key as Alania-1.

Listen Start free Docs

Try it

Hear your own Turkish.

Whatever you type is generated live. Pick a voice, tell it how to speak, or describe a new one.

70/280

One of the built-in voices.

Ready

Alania-2 · 48 kHz

AI-generated audio. Prerecorded samples are labeled during playback.

What it does

One model, four ways in.

11 built-in voices

Eleven female and male voices. F1 is the default female voice and Öncel the male one; the same timbre on every request, chosen with voice.

“Merhaba, ben sizin dijital asistanınızım. Size bugün nasıl yardımcı olabilirim?”

Style instructions

Preview

Tell it how to speak with instructions: slow and calm, whispering, more cheerful. In preview; it does not always follow the instruction closely.

“Bu konuda size yardımcı olmaktan memnuniyet duyarım.”

instructions: whispering softly

Describe a voice

Preview

A new voice from a voice_description of its age, tone and energy. No recording needed. In preview.

“Siparişiniz kargoya verildi, iki gün içinde elinizde olur.”

voice_description: A young man with a clear, energetic voice

Voice cloning

Beta

Copy a voice you have permission to use from a short recording and its transcript. Every request carries an explicit consent: true.

{ "model": "alania-v2",
  "input": "…",
  "reference_audio": "<base64 wav>",
  "reference_text": "…",
  "consent": true }

Measurements

The numbers, and how we measured them.

Sample rate: 48kHz; Twice Alania-1’s. Full-band audio.
First audio: ~190ms; Streamed request to first audio chunk, median, measured through the public API with the network included.
Character error rate: 2.7%; Our 292-sentence Turkish test set, built-in female voice; generated audio transcribed with Whisper.
Characters per request: 5,000; Long text is split at sentence boundaries and arrives as one audio stream.

API

Same endpoint. New fields.

Alania-2 is the default model on POST /v1/audio/speech. Keep your OpenAI client; add the new fields to the request body.

model: alania-v2 (default) or alania-v1
voice: alania2-f1 (default), alania2-oncel (male) or another built-in voice, such as alania2-nisa
instructions: How to speak, up to 200 characters
voice_description: A new voice described in words; replaces voice
stream: When true, audio arrives as it is generated (48 kHz PCM16 WAV)

Alania-1 keeps working: model: "alania-v1".

API reference

curl https://voice.patientdesk.ai/v1/audio/speech \
  -H "Authorization: Bearer pd_live_..." \
  -H "Content-Type: application/json" \
  -d '{ "model": "alania-v2",
        "input": "Randevunuz yarın saat 14:05 için oluşturuldu.",
        "voice": "alania2-f1",
        "instructions": "speaking slowly and calmly" }' \
  --output randevu.wav

Duyu · Speech to text