Financial services firms record and transcribe calls because regulation requires them to keep records and because a searchable archive is the only practical way to supervise it. The audio is hard: trader slang, dense numbers, company names and tickers, often over telephone-quality lines. Accuracy on figures and identifiers matters more than overall word error rate, since one misheard number can change what a trade means.
Key takeaways
- MiFID II requires investment firms to record calls and electronic communications relating to transactions and keep them for five years, or up to seven where the regulator requests.
- Transcription turns a recording archive into something compliance can search and monitor, for example for market abuse surveillance under MAR.
- The critical words are numbers, currencies, tickers, ISINs and counterparties. "Fifteen" versus "fifty" basis points is one word and a very different trade.
- Financial audio includes trader turrets, open lines and earnings calls. Public benchmarks such as Earnings-22 and SPGISpeech cover only the last of these.
- Outsourcing transcription is ICT third-party risk under DORA, so vendor due diligence, contracts and exit plans apply.
Call recording rules#
Under Article 16(7) of MiFID II (Directive 2014/65/EU), investment firms must record telephone conversations and electronic communications relating to transactions concluded when dealing on own account and to client order services. The obligation covers conversations intended to result in a transaction even if none is concluded. Records must be kept for five years and, where requested by the competent authority, up to seven. Commission Delegated Regulation (EU) 2017/565, Article 76, sets out the detail: firms need a written policy, must inform clients that calls are recorded, and must periodically monitor the records.
In the UK, the FCA's rules in SYSC 10A implement equivalent requirements. Many banks record beyond the regulatory scope, for example all front-office lines, as a matter of policy.
The Market Abuse Regulation (Regulation (EU) No 596/2014) requires firms that arrange or execute transactions to have systems to detect and report suspicious orders and transactions (Article 16). Firms commonly extend surveillance to communications, and at the volumes involved, that means transcribing calls and running searches or classifiers over the text. A recording nobody can search is of little help when a regulator asks for every conversation about a given instrument in a given week.
The language of the trading floor#
Trading conversations are fast, clipped and full of conventions that a general model has rarely heard.
| Spoken | Meaning | Typical ASR failure |
|---|---|---|
| "a yard of euros" | one billion EUR | Transcribed literally; figure lost |
| "cable" | GBP/USD exchange rate | Correct word, but no link to the currency pair |
| "fifteen bips" | 15 basis points | "fifty bips", or "fifteen bits" |
| "ten at twelve" | bid 10, offer 12 | "ten eight twelve" |
| "mine", "yours" | I buy, I sell | Correct words; meaning depends on speaker attribution |
| "A S M L" | ASML Holding ticker | "a small" |
Beyond slang, conversations carry company and fund names, analyst names, and identifiers such as ISINs (for example NL0010273215) read aloud character by character. These are structured identifiers with a fixed format, which is useful: a system told to expect twelve characters starting with two letters and ending in a check digit can recover a code that a general model would scatter into words. See numbers and alphanumerics.
Formatting matters as much as recognition. "one point two five billion dollars" should become "$1.25bn" for search to work; "twenty-five basis points" should become "25 bp". This is inverse text normalisation, and its errors are silent: the words are right, but a search for "1.25bn" misses them.
Audio sources#
- Trader turrets and dealer boards. Many simultaneous lines, open "hoot" channels broadcasting to a desk, and background floor noise. Recordings may mix several conversations.
- Client and advisory calls. Standard telephony at 8 kHz, often with two channels, one per party.
- Mobile devices. Firms must record relevant calls on business mobiles too, usually via carrier-level or app-based recording.
- Earnings calls and investor meetings. Long-form, prepared remarks followed by Q&A with analysts on phone lines.
Public benchmarks cover the last category best. Earnings-21 and Earnings-22 (Del Rio et al.) are collections of earnings calls, Earnings-22 selected to include a wide range of accents. SPGISpeech (O'Neill et al., 2021) provides about 5,000 hours of transcribed earnings-call audio with fully formatted reference text. None of these represents turret audio or trader slang, so a firm evaluating vendors should build its own test set from anonymised internal recordings.
Measuring what matters#
Aggregate WER is a starting point. For surveillance and record-keeping, add:
- Numeric accuracy: prices, quantities, rates and dates, scored after normalisation.
- Entity recall: instruments, issuers, counterparties and staff names from a firm-specific list. See entity and keyword accuracy.
- Identifier accuracy: ISINs, account numbers and trade references, exact match per identifier.
- Search recall: for a set of known conversations, does a lexicon-based surveillance query find them in the transcripts?
Speaker attribution matters because meaning depends on it: "mine" from the client and "mine" from the trader are opposite trades. Dual-channel recordings avoid most attribution errors.
Security, outsourcing and privacy#
Recorded calls contain personal data about clients and staff, and often inside information. The Digital Operational Resilience Act (Regulation (EU) 2022/2554), applicable since 17 January 2025, treats contracts with ICT third-party providers, a transcription vendor among them, as subject to risk management: due diligence, contractual provisions on data location, access and audit, and exit strategies. Banks also apply the EBA Guidelines on outsourcing arrangements where the service qualifies as outsourcing.
Where card payments are taken over the phone, PCI DSS prohibits storing sensitive authentication data such as the card verification code after authorisation. Firms typically pause recording during card capture or redact the digits from audio and transcript. Under the GDPR, clients and staff must be told calls are recorded, and the legal obligation under MiFID II supplies a lawful basis for the recording itself; secondary uses such as analytics need their own analysis. See GDPR and voice data and EU data residency.
How Reson8 approaches this#
Finance is one of the industries Reson8 serves. On the Open ASR Leaderboard short-form English sets (April 2026), Resonant-1 scored 9.02% WER on Earnings22 and 3.05% on SPGISpeech. Patterns lets a caller supply an expected identifier format, such as an ISIN or trade reference shape, and the model biases toward it; on order numbers this reduced WER from 8.0% to 1.5%. Instrument lists and counterparty names can be supplied through text-only domain adaptation. Processing runs on EU GPUs with zero audio retention and no training on customer data; Resonate Labs is ISO 27001 certified, provides a GDPR Article 28 DPA, and offers on-premise and VPC deployment on enterprise plans.
Common questions#
Does MiFID II require transcripts?
No. It requires recording and retention. Transcription is how firms make the archive usable for monitoring and for responding to regulator requests.
Can transcripts replace the recordings?
No. The regulatory record is the recording. Transcripts are derived data that help find and review it.
Should voice surveillance run in real time?
Most surveillance runs post-call in batch, which is cheaper and can use the whole conversation as context. Real-time use cases, such as prompting an adviser to read a disclosure, exist but are less common.
How do you handle multiple languages on one desk?
Traders switch between English and a local language, sometimes within a sentence. Test the system on real mixed-language calls; see code-switching.
References#
- Directive 2014/65/EU (MiFID II), Article 16(7). EUR-Lex.
- Commission Delegated Regulation (EU) 2017/565, Article 76. EUR-Lex.
- Regulation (EU) No 596/2014 (Market Abuse Regulation), Article 16. EUR-Lex.
- Regulation (EU) 2022/2554 (Digital Operational Resilience Act). EUR-Lex.
- Financial Conduct Authority. FCA Handbook, SYSC 10A: Recording telephone conversations and electronic communications.
- Del Rio, M. et al. (2022). Earnings-22: A Practical Benchmark for Accents in the Wild. arXiv:2203.15591.
- O'Neill, P. K. et al. (2021). SPGISpeech: 5,000 hours of transcribed financial audio for fully formatted end-to-end speech recognition. Interspeech. arXiv:2104.02014.