Munsit converts spoken Arabic into accurate, structured text — high-accuracy recognition across dialects and accents, with word-level timestamps, speaker diarization and meeting intelligence built on top.
1
Your first transcription
One POST with a file. Munsit analyzes the recording, converts the Arabic speech to text, and returns the transcript with its duration and word-level timestamps.
bash
curl-X POST "https://api.munsit.com/api/v1/audio/transcribe" \-H"x-api-key: $MUNSIT_API_KEY" \-F"file=@meeting.mp3"
No dialect parameter. Munsit is optimized for Arabic speech recognition with strong coverage of dialect and accent variation — the same call handles all of it. For speakers who mix Arabic and English mid-sentence, switch to the munsit-en-ar model below.
2
Choose a model
Two models. Every batch endpoint accepts an optional model parameter; if omitted, munsit is used. Live streaming takes the same model values. The transcribe endpoint and live streaming also accept an optional language: with munsit-en-ar it is auto (the default), or ar / en to set the main language of the speech; every value still handles mixed Arabic-English speech. munsit is Arabic only (ar). Custom vocabulary (hotwords) works with both models in batch; in live streaming only munsit applies it — on munsit-en-ar the hotwords you send come back in Metadata.dropped_hotwords.
Model
ID
Use it for
Default
Munsit
munsit
Arabic. Optimized for Arabic speech recognition, with strong performance across dialects and accents.
Yes
Munsit En-Ar
munsit-en-ar
Mixed Arabic–English spoken content with code-switching support — for speakers who naturally alternate between the two languages within the same conversation or utterance.
—
3
What you can send
Twelve audio formats, covering common output from different recording platforms — no transcoding step. Upload the file, or pass a public audio_url and let Munsit fetch it — see choosing a source.
Structured transcripts optimized for meeting use cases.
Live streaming
No limit
Unbounded while the connection stays active. Keep the session alive with KeepAlive during pauses; unbroken speech is force-segmented about every 60 seconds.
Longer recordings? For Audio Transcription or Minutes of Meetings, split the audio into shorter segments for best performance.
Sending more data than the model reads. Recognition runs on 16 kHz mono audio, so anything above that is downsampled on arrival. Higher sample rates, stereo channels and uncompressed WAV only add upload time. If your source is larger than it needs to be, encoding to mono 16 kHz before sending cuts transfer time with no effect on the transcript.
Working with an AI assistant? Every page is available as Markdown: add .md to the URL, or send an Accept: text/markdown header. For the whole documentation in one request, point it at llms-full.txt; the page index is llms.txt. Or use Copy Page, top right.
