Transcription
Live streaming
Section titled “Live streaming”The transcript appears in real time as you speak. Words stream into the session view while the recording is active — you don’t need to stop to see results.
Gap markers
Section titled “Gap markers”If a segment of audio fails to transcribe (network error, provider timeout, or inaudible audio), Quendian inserts a gap marker in the transcript to mark the missing span:
_[transcription failed — Xs]_The duration of the gap is shown. The recording continues uninterrupted; only that segment is missing from the transcript.
Inaudible markers
Section titled “Inaudible markers”Very quiet or unclear audio may produce an [inaudible] marker inline rather than a gap. This is normal for low-volume audio or cross-talk.
Word-level timestamps and click-to-jump
Section titled “Word-level timestamps and click-to-jump”When using Whisper-1, every word in the transcript is timestamped. Click any word to jump audio playback to that exact moment in the recording — useful for reviewing or verifying a specific phrase.
Accuracy
Section titled “Accuracy”Quendian presents transcripts verbatim — it does not post-process, spell-correct, or rewrite the provider’s output. Whatever your transcription provider returns is what you see, down to the misheard word. The accuracy of any session is therefore bounded by the provider and model you chose.
The tradeoffs differ by provider. The patterns below are qualitative observations from everyday meeting audio, not formal benchmarks:
| Provider | Strengths | Watch out for |
|---|---|---|
| Whisper-1 (OpenAI) | Word-level click-to-jump; mature baseline on clean audio | Occasional homophone confusions (“Bay Area” → “bear”); accented speakers; technical jargon |
| GPT-4o Transcribe / Mini Transcribe (OpenAI) | Cleaner output than Whisper-1; better on imperfect mics | Segment-level only — no word-level click-to-jump |
| Groq Whisper | Fastest inference; word-level click-to-jump (Whisper Large v3 model) | Occasional word swaps; same Whisper-family weaknesses as Whisper-1 |
| AssemblyAI (Universal-3-pro) | Strong on accented English and meeting audio with crosstalk; native speaker labels | Audio is processed on AssemblyAI’s servers (US or EU region selectable) |
| Deepgram | Strong on noisy meeting audio; native speaker labels; synchronous response (Nova-3 model) | Audio is processed on Deepgram’s servers (US or EU region selectable) |
| Quendian Cloud | Zero setup — no API keys, no provider sign-ups; native speaker labels | Requires a Quendian Cloud subscription; no word-level click-to-jump |
| Custom (OpenAI-compatible) | Self-hosted or local models for full audio privacy | Accuracy depends entirely on the model behind your endpoint |
For meetings where transcript quality matters most, the natively-diarizing providers — AssemblyAI, Deepgram, or Quendian Cloud — are usually the better pick. They’re tuned for meeting audio and produce speaker labels server-side, without the on-device diarization tradeoffs.
Switching providers
Section titled “Switching providers”Your transcription provider is set in Preferences → Transcription. The setting is global and applies to all future recordings. Switching does not retroactively re-transcribe past sessions.
See API Keys for setup instructions for each provider.