Voice remains one of the most resource-intensive surveillance channels. While email and messaging coverage is mature, voice introduces practical voice surveillance design challenges that are often overlooked.
These include how to handle transcription accuracy across multilingual teams, when to enable speaker separation, how to manage translated summaries and how to prioritise ingestion when surveillance coverage must be selective.
This post outlines five key issues to consider when shaping a voice surveillance approach.
One of the latest themes is the idea of only ingesting or reviewing a fixed number of minutes per call. This is sometimes proposed as a cost-saving tactic but limiting coverage, without a risk-based rationale, results in analysts receiving partial transcripts with little context. It can force them to listen to the full call to assess relevance even when key moments were never transcribed.
We will break that down, along with other core design issues, in the sections below.
Transcription Accuracy Is Still the First Barrier
While overall recognition rates have improved, error rates remain significant when calls involve non-native accents, frequent switching between languages or references to products, venues or trade-specific jargon. Background noise, poor audio setups or crosstalk make these challenges worse.
These are not new issues, but they continue to affect the quality of alerts and the confidence analysts place in them. Word-level inaccuracies often cause phrases to be missed or distorted. This often means analysts must listen to the audio themselves to verify the facts, increasing manual workload and undermining the efficiency of the transcription layer.
The problem becomes more complex when calls involve multiple languages. Some platforms rely on single-language transcription models. Others support multi-language models, but both require accurate language detection at the outset. Using the wrong model reduces transcription quality, and whilst word accuracy in these situations is not universally poor, it becomes poor when model selection does not match the language dynamics of the call.
Rather than reconstruct meaning from low-quality transcripts, most analysts fall back to reviewing audio. If transcription isn’t reliable, it no longer serves a purpose. Calls in multilingual or cross-regional environments are particularly exposed to this issue unless addressed at ingestion and processing stages. Accurate language detection shows which languages are in use and supports recruitment of surveillance analysts.
Speaker Separation Isn’t Always Worth the Effort
Speaker separation is often treated as a required step, especially in environments where calls are not recorded in stereo or lack metadata identifying participants. When a single audio track includes all speakers, the ability to segment voices appears essential, but you should asses this based on the call’s risk profile.
In one-to-one conversations or trader calls, distinguishing speakers supports accountability and traceability. In group meetings, especially those that are administrative or broad in nature, the value of speaker separation diminishes. Segmenting every person in a ten-person town hall or weekly update often creates confusion, especially when the separation logic works inconsistently.
What is meant to provide structure ends up undermining it. Anyone who has reviewed large chat channels knows how hard it is to follow the conversation when speaker attribution is scattered. Group call transcripts with inaccurate speaker segmentation are no different. They may appear thorough but often deliver misleading precision or create ambiguity that distracts from the underlying content.
It is important to recognise that surveillance regulation does not require perfect speaker attribution for all call types. Applying speaker separation to every meeting or internal update may satisfy a design checklist but does not always reflect actual surveillance need. The decision to apply it should be driven by context, not by default.
For any voice or surveillance vendors reading this, it’s worth placing more emphasis on call metadata analysis during pre-processing and detection, as this can improve how calls are risk-scored and prioritised for review.
Translation Helps Triage but Not Surveillance
Using translation to support call triage is a practical step, especially for global firms. But translation should not become the core of the surveillance process. There is a difference between making content more accessible and using that translated content to drive alerts, detection or case decisions.
Consider a Japanese-language call that’s transcribed at 85% word accuracy, then loses another 15% when translated into English. That leaves the surveillance process running on roughly 70 percent reliable data. This might be enough to support triage but not enough to support decision-making around potential misconduct or escalation.
Translation tools generally produce line-by-line equivalents of the input transcript but do not independently summarise. When translation is combined with AI-generated summaries, further loss occurs. Summaries created from translated transcripts, or created in the original language and then translated, often sound fluent but can obscure the original phrasing. Phrases of concern can be paraphrased, omitted or diluted in the process.
This has an impact when call summaries are used to assign a risk score, trigger alerts or guide escalation. If the reviewer only sees the summary and not the full transcript or audio, the opportunity to detect nuanced phrasing or intent is lost.
Surveillance should not rely on summaries of translations of transcripts. It must retain access to the source and the ability to validate language and phrasing in its original form. Translation is helpful, but it must be handled with care and used only where the risk of loss is understood and mitigated.
Fixed-Minute Limits Create Partial Visibility
As mentioned earlier, one of the current themes is to consider applying minute-based ingestion rules to reduce cost. One approach is to exclude calls that exceed a certain length. Another is to ingest but only transcribe the opening portion of a longer call. In either case, they create blind spots.
Not ingesting longer calls assumes they are lower risk and whilst this may be true in some cases, it ignores how real conversations unfold. People often take time to settle into a call. The tone may change and risk-related discussion or inappropriate behaviour often appears later, when participants are more comfortable or under time pressure. Calls excluded purely on duration may be the very ones that carry the most signals.
Ingesting but only transcribing the first segment of a long call creates different problems. Analysts will receive an incomplete transcript where if anything triggers concern in the first few minutes, they must revert to the full recording. This adds work and frustration rather than reduces effort. In many cases it creates alert fatigue where analysts start to question whether the transcription process adds any value at all.
The better approach is to assess calls based on profile, not just duration. Internal versus external, number of participants, the role of the organiser and the business area involved are more meaningful indicators of where risk may reside. All-hands meetings, HR-led calls or weekly updates are not equivalent to one-to-one conversations or trader-client interactions.
This logic applies beyond market abuse. In conduct surveillance, it is well recognised that tone and behaviour may shift during the second half of a call. People may become frustrated, animated or careless with language. Limiting ingestion or transcription to a fixed number of minutes removes visibility of precisely the type of behaviour that surveillance is intended to capture.
Ingestion models need to reflect the structure of real calls, the roles of those involved and the nature of the discussion. That means knowing the trader, understanding the desk and considering the event context.
Inclusion Should Reflect Surveillance Purpose
Surveillance of every call may seem like the safest choice but is often the least effective. Without clear boundaries, surveillance ends up covering calls that carry little relevance while consuming time and resources better spent elsewhere.
This does not mean ignoring internal meetings or control function calls. From a market abuse perspective, these calls are usually lower priority. From a conduct risk perspective, they may carry more value. People tend to be more open or less guarded in internal environments. Surveillance coverage should reflect this and adapt to the type of risk being addressed.
It is important to be deliberate, not default-driven, in deciding what to include in surveillance coverage so that it reflects actual risk. If the goal is to identify abusive behaviour, coverage should align to the communication types and participants where that behaviour is likely to surface. If the goal is to detect conduct deterioration, broader coverage may be appropriate but still needs to be prioritised.
These decisions should be based on active review, not fixed assumptions. Priorities shift as internal concerns emerge, desk structures change or external events raise the risk profile of certain users. Tagging of users, meeting types and communication patterns should be reviewed and refined over time as surveillance is a dynamic process. It cannot rely on static design.
Treating all calls the same rarely delivers consistent value for surveillance. Translated to today’s corporate vernacular, uniform inclusion rules rarely deliver uniform value. Surveillance coverage should be selectively designed to reflect real risk, which requires the surveillance function to shape upstream design decisions and align review logic with what is actually being ingested.
Conclusion : Voice Surveillance Starts With Knowing the Business
“For their risk assessments, firms should regularly take steps to consider whether their employees and/or clients may be conducting insider dealing or market manipulation. This could be achieved by transaction, order and communications surveillance, with consideration given to the employee’s or client’s usual trading behaviour and/or strategies…”
— FCA Financial Crime Guide, Release 49, July 2025
Many of the challenges in voice surveillance stem from decisions being made too far downstream. The tools may be in place but if ingestion is driven by default rules, time limits or blanket inclusion, the system ends up generating noise and adds unnecessary time to the review process. Transcription quality, speaker separation and translation tools can support surveillance but do not replace the need for intentional design.
Designing voice surveillance coverage often depends on understanding how risk actually appears in real conversations. In many environments it means building familiarity with how trading desks operate, how different call types unfold and where the line sits between routine communication and sensitive content. These are the practical elements that help shape meaningful coverage…trader, trade flow and timing.
When upstream design reflects how business is done, surveillance becomes more targeted and more efficient. Tagging of users, meeting types or topics can be refined over time and prioritisation adjusted as patterns emerge. Ingestion can be selective without being blind and review teams can spend more time on relevant cases, not compensating for gaps in the data.
For firms reassessing their voice surveillance approach, the next step is unlikely to be more data. It is making the design of surveillance coverage more deliberate by profiling conversations based on who is speaking, what is being discussed and when risk tends to emerge.
The same design principles used in trade surveillance apply here. Trade surveillance is often built around the trader, their trade flow and the timing of their activity. Voice calls are part of that same trading footprint as they sit alongside the trades, as part of the overall behaviour, so the design of voice surveillance should follow the same logic…centred on the trader, structured around the flow and sensitive to timing.
We work with firms to design upstream models grounded in real desk activity, applying a structured Trader, Trade Flow, Timestamps approach to help prioritise voice ingestion and align it to risk.
If this commentary resonates with the challenges you’re facing and you’d value some input on how to approach them, feel free to get in touch.
