Speech-to-text built for the trading floor

A bond salesperson who says “I’ll pay fifteen for a yard” has made an exact statement. The price is fifteen after leading digits both sides already know, and a yard is a billion. A general speech recognizer hears the same sounds and tends to write something about gardening. That distance between what a trading floor says and what a consumer engine expects to hear is why speech-to-text for the markets has become a specialty of its own.

What general engines expect to hear

A recognizer makes two guesses at the same time. An acoustic model estimates which sounds were spoken, and a language model estimates which sequences of words are likely. In noisy audio the second guess carries much of the weight, and it is learned from whatever text the engine was trained on: dictation, search queries, broadcast news, call center conversations. Nothing in that material teaches it that “cable” usually means sterling against the dollar, that “Swissy” is a currency, or that “mine” called out across a desk closes a purchase.

Trader speech is also full of numbers said in shorthand. “Two and a half, three and a quarter” is a two-way price. “Fifteen, twenty” may be a bid and an offer with the big figure left off because nobody on the line needs it. Tickers and bond names come out as clipped syllables or strings of letters. A general model that drops the fraction from a price, or hears “at” where the trader said “by”, has written a tidy sentence and a wrong record.

The audio is harder as well

Published results for conversational recognition come mostly from the Switchboard telephone corpus, in which strangers chat about an assigned topic. On that test the numbers have fallen fast. In October 2016 one research team reported a system that matched the 5.9 percent error rate it measured for professional transcribers. In March 2017 another group reported 5.5 percent for its own system, and measured the best human transcriber at 5.1 percent after quality checks. Even the human benchmark is still in dispute.

A trading desk sounds nothing like Switchboard. Turret handsets and speaker lines carry several conversations at once. People talk over one another, and the loudest voice may belong to someone not on the call. Telephone audio cuts away the higher frequencies that help separate an “f” from an “s”. A system that rivals human transcribers on a clean research set can do much worse at eight in the morning on a rates desk.

What a finance-trained engine changes

The remedy is to train on the domain. Engines built for markets feed their language models with transcripts of trading conversations, so that “yard” after a number becomes the likely reading and “cable” stops being a wire. Pronunciations for tickers, currency nicknames and product names go into the lexicon. Acoustic models learn from turret and handset recordings instead of studio speech. The output is then normalized into the form a desk would write, so that “one oh five and a quarter” comes out as 105.25.

Errors remain, but training on the domain moves them away from the words that matter. A transcript that misspells a name during small talk costs nothing. One that turns fifteen into fifty in the size of an order causes trouble for everyone who reads it later.

A person still checks the text

A sensible deployment keeps a person in the loop. A trader or an assistant reviews the transcript of an order and fixes what is wrong, and those corrections go back into training so the engine picks up that desk’s vocabulary over time. Another approach uses the transcript only to find the right minute of audio, which someone then listens to before relying on it. Either way, the recording remains the record, and the text is an index to it.

Why this matters before January 2018

The European MiFID II directive applies from 3 January 2018. Its Article 16(7) requires investment firms to record telephone conversations and electronic communications relating to transactions concluded when dealing on own account and to client order services. The duty covers calls that are intended to result in a transaction even where no trade follows, and the records must be kept for five years, or up to seven where the regulator asks. Clients have to be told beforehand that their calls will be recorded.

Recording on that scale leaves years of audio on every trader. Storing it is cheap; finding one conversation when a regulator or a client asks for it is not. Searchable text, even imperfect text, turns hours of listening into a list of hits to check. In the United States, CFTC Rule 1.35 already requires many futures firms to keep oral communications that lead to a trade, and desks there face the same volume problem.

Testing on real desk audio

A figure quoted from a public benchmark says little about a particular floor. The useful test is local: take a sample of a desk’s own recordings, have them transcribed by hand, and score the machine against that. It helps to count errors on sizes, prices, directions and instrument names separately from the overall error rate, because those are the words a compliance officer or a client will actually read.