Transcription

Transcription runs on the chair's device, which holds every recording.

Language

Set Language spoken to the language of the speeches: English or Bangla (বাংলা).

English works offline and with every cloud provider.
Bangla needs a cloud provider with a key, because the offline models only understand English. In Auto mode Bangla always goes to the cloud and never falls back to the offline model; Offline only can't transcribe it.
Use ElevenLabs Scribe for Bangla. On 20 public Bangla recordings it got 15% of words wrong, against 82% for Whisper large-v3 on Groq, and most of its "errors" were English words written in English, which suits debates where speakers switch languages.

Modes

Home → Transcription settings:

Mode	What happens
Auto (default)	Cloud if a key is set and you're online, otherwise offline. If the cloud request fails, it falls back to offline.
Offline only	Audio never leaves the device.
Cloud only	Fastest and most accurate; needs internet and sends audio to the provider.

Offline models

English-only Whisper models, downloaded once (do it on Wi-Fi before an event):

Model	Android	Windows	Notes
Tiny	31 MB	40 MB	Fastest, least accurate
Base	57 MB	76 MB	Recommended
Small	181 MB	250 MB	Most accurate, slow on phones

On a Windows PC, Base transcribes roughly 4× faster than real time (a 7-minute speech in under 2 minutes). Phones are slower; keep the app open while it works, as the screen stays awake during transcription.

Cloud providers

One-tap presets:

ElevenLabs Scribe: https://api.elevenlabs.io/v1, model scribe_v2. Best for Bangla. Key from elevenlabs.io → Settings → API keys.
OpenAI: https://api.openai.com/v1, model whisper-1
Groq: https://api.groq.com/openai/v1, model whisper-large-v3-turbo. Free tier, good for English.

Any other OpenAI-compatible /audio/transcriptions API works too: pick OpenAI or Groq, then change the URL and model.

Paste the API key and Save. The key is stored on the device unencrypted, and audio is sent to the provider you choose.

To try a provider on a recording before a tournament, contributors can use npm run transcribe:try with keys in .env (see CONTRIBUTING).

Using transcripts

On a round's review screen: Transcribe per recording or Transcribe all. Jobs run one at a time. Tap any line to play the audio from that moment; the line playing is highlighted. Transcripts can be included in the PDF report.