01 · Accuracy
Heard right the first time.
Lowest WER on three Turkish sets: conversation, read speech and call center. A vocabulary hint (prompt) is supported for drug names, doses and ID numbers.
Try it
Speak, then read.
Turn on your microphone and say a sentence. A draft appears while you speak; the final text arrives when the sentence ends.
Read this aloud
“Yarın saat 14:30'daki diş randevumu 16:00'ya alabilir miyiz?”
Checking availability
Ready to listen
Transcript
Press the microphone and speak Turkish.
Measured on three Turkish test sets.
Same recordings, same normalization, the same measurement for every model.
voicedata-TR · 235 multi-speaker Turkish clips
- Duyu8.54
- whisper-large-v310.97
- whisper-large-v3-turbo11.19
- Qwen3-ASR-1.7B13.65
- mihuai/turkish-stt19.89
- Qwen3-ASR-0.6B22.30
Word error rate (%) · Turkish-normalized · lower is better
Built for real conversations.
02 · Turn taking
Knows when you start and finish.
In-model voice activity detection (vad), segment_start / segment_end events and a final after 500 ms of silence. No external VAD.
03 · Telephony
Works at 8 kHz too.
Trained on telephone-band data; server-side resampling. 9.9 WER on the sierra-tr call-center set.
04 · Formats
Text, timing, subtitles.
json, text, verbose_json, srt, vtt; word-level timing via timestamp_granularities[]. Drop-in compatible with OpenAI Whisper clients.
You stop speaking, the text is ready.
- End of speech to final text
- ~120ms
- Live stream · p50
- Faster than real time on files
- 50×
- Audio duration ÷ processing time · up to
API
Connect with the client you have.
An OpenAI-compatible /v1/audio/transcriptions for files, WebSocket events for live audio.
Output formats
- json
- text
- text
- plain text
- verbose_json
- segment and word timing
- srt
- subtitles
- vtt
- web subtitles
curl https://voice.patientdesk.ai/v1/audio/transcriptions \
-H "Authorization: Bearer pd_live_..." \
-F file=@kayit.m4a \
-F response_format=json