Our work / VOX
VOX
Nepali speech-to-text platform
Converts spoken Nepali into written text for professional and official use, with speaker separation and noise reduction. Trained on custom Nepali speech corpora spanning regions, age groups and everyday contexts.
- Category
- Speech recognition
- Platform
- Web
- Accuracy
- 95% transcription accuracy, 15% character error rate
- Models
- Wav2Vec 2.0 (acoustic)Pyannote (diarization)
- Trained with
- PyTorch
What the product does
Capabilities
High transcription accuracy
95% accuracy on Nepali speech-to-text conversion, giving precise and reliable transcription for clear speech.
Multi-speaker recognition
Speaker diarization distinguishes between voices, which is what makes VOX usable for meetings and discussions rather than monologues.
Localized for Nepali
Tailored to the Nepali language and its dialects, so native speakers get transcription accuracy and usability that generic tools do not offer.
Optimized for professional use
A 15% character error rate together with noise reduction keeps performance reliable in government and professional settings.
Continuous learning
Adaptive learning and ongoing model refinement mean accuracy and reliability keep improving after deployment.
Built for real rooms
Noise reduction algorithms are part of the pipeline, not an afterthought, for recordings made in imperfect conditions.
How it is put together
Architecture
- Acoustic model
- Wav2Vec 2.0, fine-tuned to minimise transcription error on Nepali speech.
- Speaker separation
- Pyannote diarization to tell voices apart across a recording.
- Context handling
- Acoustic models tuned to capture the contextual nuances a literal transcript would lose.
- Noise
- Built-in reduction algorithms for performance in challenging environments.
- Training stack
- PyTorch, with continuous adaptation and model refinement.
Who it is for
Where it is used
Government and public record
Accurate documentation of proceedings and public sessions, which is what makes transparent record-keeping practical at scale.
Education and research
Lectures, seminars and interviews digitized and archived instead of lost to a recording nobody has time to type up.
Business operations
Meetings, customer support and media transcription at volume, in the language the room is actually speaking.
Built in the same workshop
