Lab 08
Voice Intelligence
See speech become words
Turn a sample recording or your own speech into captions on your device.
Overview
Spoken information is hard to search, skim or use as captions.
How this could help a business
Speech-to-text can support captions, searchable notes and review tools. A business version needs evaluation of its languages, accents, background noise and consent requirements.
What I built
- A local audio application connecting microphone capture, speech detection and streamed transcription.
- An interface for confirmed and provisional words, with confidence and alternatives available for inspection.
Live demo
Try a sample recording first, or use your microphone when you are ready.
What I built
Implementation, decisions and evidence.
What Sebastian built
The application around the technology.
- A local audio application connecting microphone capture, speech detection and streamed transcription.
- An interface for confirmed and provisional words, with confidence and alternatives available for inspection.
- Local file processing, optional text-based mood analysis and a reference-audio evaluation pipeline.
One engineering decision
Show words that may still change differently from confirmed words, so a fast partial result is not mistaken for a final transcript.
What to check yourself
Compare the sample recording with its real transcript. Use Inspect pipeline for word alternatives, and What I built for the measured evaluation.
Tools and model credits. Whisper transcribes speech, Silero detects speech activity and GoEmotions labels the emotion in the transcribed text. These are pretrained models integrated into the application.
These are portfolio demonstrations. Explore professional experience separately, or discuss how the approach could fit your team.
Learn
Understand what the application is doing.
Listen to the sample, transcribe it, and compare the words you hear with the captions. You can try your own audio afterwards.
Pause and predict
If a transcript sounds fluent, should you assume names and amounts are correct?
A learning question, not a score. Try this in the working demo.
Try changing something
Now make it your experiment.
Compare a clear recording with one containing noise. Review names and numbers especially carefully.
Transcripts can contain mistakes, and mood labels read the words, not the voice. Word error rate is measured on a 200-clip benchmark and is still above my target.
A few words, explained
- Provisional words
- An early guess that can change as more speech arrives.
Data handling & current limitations
Audio stays on your device. Optional mood analysis reads your words, not your tone of voice.
Demo only. Each lab below explains which inputs stay on your device and which data is sent for server-backed features. Model files are downloaded and may remain in your browser cache. Hosting and security services also receive the network requests needed to serve the site.
Audio and transcripts are processed in this browser and are not uploaded. Downloaded speech and optional emotion models may remain in your browser cache.
Transcripts can contain mistakes, and mood labels read the words, not the voice. Word error rate is measured on a 200-clip benchmark and is still above my target.