Why we built TakeNote on MAI-Transcribe-1

Transcription accuracy is a compliance issue

When an IFA tells a client their recommended portfolio carries a moderate risk profile, that phrase needs to appear in the transcript exactly as spoken — not paraphrased, not truncated, not misattributed to the client. In regulated financial advice, a transcription error is not an inconvenience. It is a compliance failure.

When we evaluated transcription models for TakeNote, we started with that constraint. Every other consideration — cost, latency, language coverage — came second. The model had to be demonstrably, measurably more accurate than alternatives, with particular strength in financial vocabulary, UK accents, and multi-speaker scenarios.

MAI-Transcribe-1 was the clear answer.

What makes MAI-Transcribe-1 different

Microsoft's MAI-Transcribe-1 was trained on a significantly larger and more diverse dataset than competing models, with particular depth in professional and domain-specific English. In third-party benchmarks, it consistently achieves the lowest word error rate across financial services vocabulary — terms like “suitability assessment,” “drawdown,” “capacity for loss,” and “Consumer Duty” are transcribed reliably rather than approximated.

Beyond raw accuracy, MAI-Transcribe-1 has three properties that matter specifically to TakeNote's use case:

• Speaker diarisation at scale. The model accurately separates speaker turns even in overlapping dialogue, background noise, or telephone-quality audio — the conditions of a real client meeting, not a studio recording.
• UK English optimisation. Regional accents from Scotland, the Midlands, and Wales are handled materially better than in US-trained models. This matters when your users are advisers working across the UK.
• Azure-native deployment. MAI-Transcribe-1 runs entirely within Microsoft Azure, which means TakeNote can process audio on Azure UK South infrastructure without any data leaving the United Kingdom.

The data residency requirement was non-negotiable

Before we wrote a single line of product code, we set one absolute constraint: all client data must stay within the United Kingdom. MiFID II, COBS, and the FCA's Principles for Businesses all create an environment where data sovereignty is not a preference — it is a regulatory expectation.

Many transcription providers — including some large consumer AI platforms — route audio through US or EU infrastructure by default, with UK residency available only as a paid enterprise add-on, if at all. MAI-Transcribe-1 on Azure UK South gave us UK-only processing from day one, without compromise.

Accuracy figures

MAI-Transcribe-1 achieves 3.0% word error rate on the Artificial Analysis AA-WER benchmark, ranking 4th globally and first among real-time models. Across 25 languages, it achieves the lowest average WER of any available model. This is the kind of reliability regulated firms demand.

Generic models fail in real-world conditions. TakeNote uses a model engineered specifically for the conditions financial advisers actually face: multi-speaker meetings, background noise, technical vocabulary, and the regulatory requirement for audit-proof records.

What this means for your firm

When you upload a meeting recording to TakeNote, you are not relying on a general-purpose AI tool adapted for compliance. You are using a purpose-built system — with the most accurate transcription model available — designed around the specific vocabulary, regulatory obligations, and data requirements of UK regulated financial advice.

That is a meaningful difference when your compliance record is reviewed by the FCA.