Transcription quality at scale for financial research

The problem

Financial research is sold on what was said in calls with experts and companies. In a transcript, a wrong name, number or speaker becomes a wrong fact in somebody's research. Volumes kept growing and turnaround had to stay short: human transcription alone could not keep pace, and raw speech recognition was not accurate enough to sell.

What we built

We built the pipeline and the tools the people in it use. Candidate speech engines are benchmarked against each other, and every clean-up step is tested for words the model made up. A language-model pass tidies the draft, names and terms are checked against outside sources, and speakers are attributed from the call's own context. Human editors finish each file in a purpose-built editor while automated quality checks route doubtful files to reviewers. Live calls stream through the same service, with a fallback provider if one fails mid-call.

What production made hard

• Speech models invent words when the audio is unclear, so every clean-up step was tested for words that were never said.
• A live call cannot be run again: when a provider failed mid-call, the stream had to move to another without losing the transcript.
• Speaker labels matter as much as words, because a quote given to the wrong person is a wrong fact.

What a similar project needs

• A test set of your own audio with corrected transcripts, so each model change is measured before it ships.
• Editors in the design from day one: the pipeline is built around where people add the most accuracy.
• Clear rules for what may be checked against outside sources and what must stay private.