QUARKEONLABS / AI ENGINEERING
← All projects

Speech / Evaluated demonstration

Conversation Lens

Turn conversations into reviewable actions.

Real demo recording · English narration and captions · Fictional data

THE PROBLEM

Reviewing recorded conversations and turning them into follow-up work requires replaying audio and reconstructing its context.

THE WORKFLOW

Upload fictional English speech, review timestamped segments, return to the original audio and inspect the summary and actions.

  1. 01English recording
  2. 02Local transcription
  3. 03Summary and actions
  4. 04Timestamped review

RECORDED EVIDENCE

Results, with context.

Results on synthetic data
CheckCases
Expected facts found12/12
No prohibited phrase12/12
Dataset
synthetic-regression-v3-english
Completed cases
12/12
Median / p95 per case
1.911 s / 12.658 s
Evaluation record (UTC)
2026-10-08

Source: the project’s evaluation record, using local inference. Recorded hardware: NVIDIA GeForce RTX 5060 Ti, 595.84, 16311 MiB. Latency includes the full HTTP case and may include initial model loading. These measurements have not been rerun on the machine serving this website.

Case study (PDF · EN) ↓Evaluation data (JSON) ↓Video (MP4) ↓

WHAT WE COULD BUILD WITH YOU

An audio transcription and conversation-review pilot with traceability to the source.

What we would define first

Languages, noise, consent, data handling, speaker attribution and transcription targets.

Demonstration limits

The 1.68% word error rate was measured on synthetic English speech: 4 edits across 238 words. This is not a real-call or diarization benchmark.

Discuss your use case ↗
Screenshot of the interface for Conversation Lens
Demonstration review interface.

Explore other projects