For business
What it is
Turning raw material into text: speech becomes a transcript that shows who said what, scans become readable text.
Why it matters
Everything downstream reads text. Speaker labels matter: “the client wants to retire at 60” is a different fact from the adviser suggesting it.
Risk without it
Poor transcription quietly poisons every later step: a summary can only be as accurate as the words it was given.
For engineers
How it works
Audio goes through speech-to-text with diarisation (speaker separation) and word timestamps. PDFs use the text layer when there is one and OCR only for scanned pages. Long recordings are split, transcribed in parallel and stitched back on timestamps.
Typical stack
- Whisper-class STT
- AWS Transcribe / Deepgram
- Textract / Tesseract
- ffmpeg
Pitfalls & trade-offs
- Diarisation struggles with cross-talk, so keep confidence scores.
- Domain vocabulary (ISA, SIPP, fund names) needs custom vocabulary or post-correction.
- Measure word error rate on your own recordings, not vendor benchmarks.
In this example
In
→ meeting_0412.m4a 2 speakers
Out
[00:42] ADVISER: What are your goals? [00:47] CLIENT: Retire at 60…
Fictional sample · not a client system