What transcription changed for trade surveillance
For years, a bank's recorded calls were an archive to search after something went wrong. Machine transcription turned them into something a surveillance team can read every day, which brought its own problems.

When the currency benchmark cases broke, the evidence that mattered was written. In November 2014 the CFTC fined five banks a combined $1.4 billion over attempts to manipulate foreign exchange benchmark rates between 2009 and 2012, and its orders described traders using private chat rooms to share confidential client order information and coordinate their positions. Investigators could read those rooms line by line.
Phone calls were a different matter. Firms had recorded their trading lines for years, and after MiFID II took effect in January 2018 European investment firms had to record calls intended to lead to transactions and keep them for five years. The recordings existed. What was missing was any practical way to review them before someone had a reason to look.
Text was surveilled, voice was sampled
Communications surveillance grew up around text. Email and chat can be searched as they arrive, so compliance teams built lexicons of risky phrases (“keep this quiet”, “take it offline”, slang for a benchmark fixing window) and ran every message through them. Hits went to a reviewer, who closed most of them and escalated a few.
Voice did not fit that workflow. A reviewer can read a day of one trader’s chat in minutes, but listening to a day of that trader’s calls takes a day. So most firms reviewed voice by sampling a small share of calls, chosen at random or by risk, and searched the rest only when an investigation or a complaint demanded it. Calls that nobody pulled were never heard.
What changed with transcription
Speech recognition good enough for conversational audio made it possible to turn recorded calls into text within minutes of the call ending. Once that text existed, it could flow into the same review system as chat and email. The lexicon that flagged a phrase in a chat room could flag the same phrase said aloud. A reviewer looking at an alert about a trader could see the calls, chats and orders from the same hour side by side.
This matters for the obligation that sits under all surveillance in Europe. Article 16(2) of the Market Abuse Regulation requires any person professionally arranging or executing transactions to maintain effective arrangements, systems and procedures to detect and report suspicious orders and transactions. An unusual order is easier to judge when the conversation that led to it is on the same screen.
Transcription also changed investigations after the fact. Searching a year of audio for a client’s name or a bond’s nickname used to mean listening to many hours of tape. With transcripts, it takes a query, followed by a person listening to the handful of calls that come back.
The limits
The transcripts are only as good as the recognizer, and trading audio is hard. A system that turns “fifty” into “fifteen” will mislead a reviewer checking an order size, and a lexicon cannot flag a phrase the engine never wrote down. Accents, crosstalk on open lines and desk shorthand all push error rates up, so the calls most in need of review can be the ones transcribed worst.
Volume moves the bottleneck rather than removing it. Running every call through a lexicon produces a flood of alerts, and most are harmless: a trader saying “let’s take it offline” about a lunch booking. Firms tune lexicons, add models that score the context around a hit, and still find that reviewers close far more alerts than they escalate.
Language is another gap. Many desks speak more than one, and traders switch mid-sentence. A lexicon written in English misses a warning said in another language, and recognizers are weaker in some languages than others.
People adapt as well. Traders who know their calls are transcribed learn to speak in ways a lexicon does not catch, the same way chat users learned years ago. And the easiest way around any recorded channel is a personal phone. U.S. regulators have spent the years since December 2021 fining firms for business conducted on unapproved messaging apps, starting with a $125 million SEC penalty and a $75 million CFTC penalty against a single bank.
The record stays the recording
One rule has not changed. The transcript is a tool for finding and reading calls; the audio is the record. Under the rules implementing MiFID II, recordings must be kept in a durable medium that can be replayed and cannot be altered, and a reviewer about to escalate a call normally listens to it first. An automatic transcript can point to the right thirty seconds. Whether those thirty seconds show misconduct is still a judgment made by someone who heard them.