Long-form transcription
Hours of audio split, transcribed and assembled into a clean right-to-left Word document, sent by email.
All work
Case study · Voice & speech
Arabic speech, turned into text and feedback.
Two products on one speech engine: a service that transcribes long Arabic recordings into ready-to-edit Word documents, and a practice app for schools where pupils read aloud and get instant, word-by-word feedback.
- Sector
- Education & media
- Our role
- Speech models, apps, GPU hosting
- Status
- In production
- GPUspeech recognition on our own servers
- .docxright-to-left Word documents, delivered by email
- 1 tapto record and get feedback on every word
The challenge
Off-the-shelf tools struggle with Arabic speech
Long recordings, classical pronunciation, many voices: general-purpose apps cut them short, mangle them, or return text nobody can use without retyping.
Teachers needed something no transcription service offers: feedback a child can act on — which word was missed, which was mispronounced — without showing them raw machine output.
What we built
One speech engine, two products
Read-aloud practice app
Pupils record themselves on Android; the app compares their reading with the expected text.
Word-by-word feedback
Mistakes are highlighted by severity, so a child sees exactly what to fix.
Models trained for the job
Speech models fine-tuned on our own GPU servers for the voices and texts that matter here.
Faithful audio
Recordings are captured as they are, without phone 'enhancements' that hide pronunciation errors.
On our infrastructure
Audio is processed on our own servers.
How it works
From a voice to something you can use
-
Record
Audio is captured faithfully, in the app or by upload.
-
Recognise
Our speech models transcribe it on our own GPUs.
-
Compare
The transcript is aligned with the expected text, word by word.
-
Deliver
A Word document by email, or instant feedback in the app.
Results
No more retyping, and feedback the moment a pupil stops reading
Client details are kept confidential. Full references and a live demo are available on request.
Built with
- Python
- PyTorch
- Whisper
- wav2vec2
- FastAPI
- Android
- nginx
- Long recordings transcribed without manual typing
- Pupils get feedback as soon as they finish reading
- Teachers spend their time on the mistakes that matter
- Audio stays on our infrastructure
Want something like this?
Tell us what you have in mind. We'll suggest the simplest version that does the job, with a clear quote.
Start a project