HM Courts & Tribunals Service: BenchNotes
This system enables First‑Tier Tribunal Immigration and Asylum Chamber (IAC) judges to securely transcribe their oral decisions in real time using Azure Speech Services ASR models.
1. Summary
1 - Name
BenchNotes
2 - Description
This system will be used by First-Tier tribunal IAC judges to dictate judgments after the Hearing has taken place.
The system will take spoken-word audio and transcribe in real-time. Sensitive personal information on the appellant is often used when writing up decisions.
The underlying technology used to transcribe the audio is an AI automated speech recognition (ASR) service (Azure Speech Services) which use AI models that are copies of base models, retained within the HMCTS secure infrastructure boundary and not shared for public use. The models used will not be trained using personal data.
3 - Website URL
N/A
4 - Contact email
JudicialTranscribe@justice.go.uk until mid Sep BenchNotes@justice.gov.uk from mid Sep onwards
Tier 2 - Owner and Responsibility
1.1 - Organisation or department
HM Courts & Tribunals Service - Immigration and Asylum Chamber
1.2 - Team
DTS Innovation
1.3 - Senior responsible owner
Service Owner (Tribunals)
1.4 - Third party involvement
No
1.4.1 - Third party
N/A
1.4.2 - Companies House Number
N/A
1.4.3 - Third party role
N/A
1.4.4 - Procurement procedure type
N/A
1.4.5 - Third party data access terms
N/A
Tier 2 - Description and Rationale
2.1 - Detailed description
This tool supports the writing up of judicial decisions by producing structured transcripts of dictated decisions. It is used post Immigration and Asylum Chamber (IAC) proceedings to capture judicial decisions and reasonings supporting the creation of Decisions and Reasons documents. Audio is captured via direct input (e.g. a judge’s laptop microphone). This audio is transcribed in real time, supporting judges as they draft their judgments. The Decisions form part of the evidential and procedural record used in judicial determinations. During the judgment-writing process, judges enter case information and use the dictation function to populate structured templates. This enables the creation of formal judgment documents. Judges retain oversight of this process and have the opportunity to review and ensure that the transcript accurately reflects their intended reasoning before finalising the judgment. This review step is important to maintain accuracy and judicial integrity. Once created, dictated transcripts are stored in a central repository. Judgment transcripts are embedded within templated documents, which can be downloaded and subsequently uploaded into the relevant case management system. These outputs directly inform the production of Decisions and Reasons documents, forming part of the official record that underpins case outcomes. After storage, transcripts are managed in line with the IAC Records Retention and Disposal Schedule (RRDS). This ensures that data is retained only as long as necessary and handled in accordance with governance, legal, and compliance requirements. Access to this tool is restricted to authorised users (including judges), with authentication managed through HMCTS identity services. Users can only access their own transcripts, ensuring appropriate confidentiality and control. Transcript data is not shared externally by the system. In summary, this tool supports judicial workflows by enabling the capture, review, and structuring of decisions, which directly feed into the creation of formal judgments and decisions, while ensuring appropriate governance and security throughout the lifecycle of the data.
2.2 - Benefits
The system streamlines the judgement writing process by enabling fast, accurate, secure creation and storage of judgments using dictation reducing manual effort and ensuring compliant, controlled access to sensitive information.
2.3 - Previous process
The original process involves the Judge writing up their decision post Hearing either using commercially available dictation software or typing up their notes into a Decision & Reason document template which they create and format themselves. They can then save it and manually upload the final version to myHMCTS for official documentation.
2.4 - Alternatives considered
N/A
Tier 2 - Deployment Context
3.1 - Integration into broader operational process
Real-time dictated speech is streamed directly from a judge’s laptop microphone into the system through the UI, enabling immediate transcription of their judgment.
Once the audio is ingested, it is processed by the platform’s Transcription Service. This service supports real-time transcription using pre-configured models from a large language model (LLM) repository. The user will then enter case information in a form and then use the real-time transcription to create the judgments. The output is displayed directly in the browser as the audio is processed.
For judgments, the case metadata and real-time transcript is inserted into the selected document template to produce a structured, downloadable document. All transcripts are stored in a central Transcript Repository. Judgment transcripts are saved during the real-time dictation process. These documents are subject to retention and deletion policies defined by the IAC Records Retention and Disposal Schedule (RRDS), ensuring compliance with data governance standards.
The final output of the process is a templated judgment document, generated from the real-time transcript and made available for download via the UI. This document is then uploaded by the judge into the case management platform. Access to the system is strictly controlled. Users, including judges and services, authenticate via the HMCTS Authentication Service as guest users linked to eJudiciary and Justice tenants. Only users within the approved user group can access the system, and each authorised user can only view their own transcripts. This ensures that sensitive data is protected and only accessible to those with the appropriate permissions.
Data is not shared by the system with external parties.
3.2 - Human review
Always, the tool supplements existing manual processes. The output is always reviewed and validated by the user.
3.3 - Frequency and scale of usage
Total potential userbase of 430 judges. Expecting that decisions may take an hour to dictate and carried out rougly 1 or 2 times per week per user.
3.4 - Required training
Training provided through Judicial Office to the judiciary.
3.5 - Appeals and review
Tool does not make decisions.
Tier 2 - Tool Specification
4.1.1 - System architecture
This system is a cloud-based platform designed to streamline judicial workflows by recording and transcribing judicial decisions into official court record documents. It accepts real-time audio streamed from a judge’s device, alongside user-entered case metadata and predefined document templates. The core processing layer uses AI-driven transcription services to produce live transcripts (displayed instantly in the browser). These transcripts are then combined with templates to automatically generate formatted judgment documents.
All transcripts, metadata, and outputs are stored in a central repository with defined retention policies, while observability services capture performance metrics and logs for monitoring and evaluation. Access is securely managed via HMCTS authentication, ensuring users can only view their own work, while administrators have broader oversight and configuration capabilities.
4.1.2 - System-level input
Audio stream from the host machine microphone. Text entry in form.
4.1.3 - System-level output
Text based transcript of the audio Decisions & Reasons legal document.
4.1.4 - Maintenance
Model will not be trained, base model from Microsoft for the Speech-to-Text function which is automatically updated.
4.1.5 - Models
Universal Language Model: https://learn.microsoft.com/en-us/azure/communication-services/concepts/call-automation/real-time-transcription
Tier 2 - Model Specification
4.2.1. - Model name
Universal Language Model [Microsoft Azure Communication Services] - version 1.47.0
4.2.2 - Model version
No version number is published. It is always ‘latest’.
4.2.3 - Model task
The model is used by the Microsoft Real-Time Speech Service to transcribe audio inputs.
4.2.4 - Model input
Audio data (either by stream or file) is consumed by the model.
4.2.5 - Model output
Transcribed output in JSON format
4.2.6 - Model architecture
Azure uses default, multilingual STT models designed for broad, everyday speech‑recognition tasks.
The model transcribes spoken audio, both streaming and prerecorded, into readable text. It supports real‑time transcription, fast file-based transcription, and large-scale batch transcription.
4.2.7 - Model performance
A mixture of qualititve and quantitive evaluation is undertaken with the judiciary users. We seek feedback from the judiciary on accuracy metrics. We also are developing an automated accuracy testing framework that can be run repeatedly upon build and periodically overnight to continuously test the model performance against ‘known accurate’ transcripts - this will work off the ‘Word Error Rate’ metric.
4.2.8 - Datasets and their purposes
N/A
2.4.3. Development Data
4.3.1 - Development data description
N/A
4.3.2 - Data modality
N/A
4.3.3 - Data quantities
N/A
4.3.4 - Sensitive attributes
N/A
4.3.5 - Data completeness and representativeness
N/A
4.3.6 - Data cleaning
N/A
4.3.7 - Data collection
N/A
4.3.8 - Data access and storage
N/A
4.3.9 - Data sharing agreements
N/A
Tier 2 - Operational Data Specification
4.4.1 - Data sources
User inputs, microphone audio input stream.
4.4.2 - Sensitive attributes
Personal data (within audio data) is handled within the software solution, which includes processing and storing of audio files and transcripts. Data retention policies apply.
Given the nature of the data (Immigration and Asylum Chamber decisions), the data will often include personal background information of individuals, including sensitive information that could be factors in making an asylum claim (criminal data and special category data).
In addition; Case metadata will be collected and stored for the purposes of constructing the decisions & reasons document in conjunction with transcript content.
4.4.3 - Data processing methods
Pre-processing on the audio file to improve audio quality (normalisation, echo) - controlled through FFMpeg.
4.4.4 - Data access and storage
Transcripts and generated documents are stored within the storage services in Azure - and are only accessible by the user which generated them. Whilst technically the information stored isn’t necessarily sensitive to other users of the system, we have decided to keep permissions simple. All data is encrypted at rest in the Microsoft Storage Account.
4.4.5 - Data sharing agreements
N/A
Tier 2 - Risks, Mitigations and Impact Assessments
5.1 - Impact assessments
The following impact assessments have been completed at various points 03/2026:
- Data Protection Impact Assessment
- Human Rights Impact Assessment
- Equality Impact Assessment
- Ethics Impact Assessment
- Responsible AI Assessment
- Staff Impact Assessment
There are a number of rights and risks that have been identifeid that relate to:
- rights and risks of children, vulernable groups, and fairness of trial.
Mitigations include:
- manual (human in the loop) requirement for all judicial processes, accuracy testing, data safeguards for privacy and security.
Overall Consequential Impact Assessment:
The use of AI supported transcription in IAC judgments presents a manageable, medium level impact mainly due to the public interest context and potential scrutiny around introducing AI into judicial processes. Clear communication about the distinction between generative AI and automated transcription, and the judiciary’s continued responsibility for accuracy, helps mitigate risk.
Operational effects are limited to workflow adjustments for court clerks and increased drafting efficiencies for judges. The system is technically low risk, relying only on existing HMCTS authentication, with no additional cost exposure and potential efficiency savings. Overall, impacts are operational and reputational rather than financial or structural and can be effectively managed through transparent explanation and standard governance.
If the solution fails to meet its objectives, the consequential impact is limited, as existing manual and non‑AI transcription processes can continue unchanged. The tool supplements rather than replaces current practice, ensuring continuity of service.
Overall Complexity Impact Assessment:
The project operates within a single jurisdiction and requires no political engagement, with cybersecurity and data‑protection considerations already embedded through a completed DPIA and established governance processes. It introduces new technology into HMCTS but builds on an existing, productionised MoJ platform. UK‑only data residency, restricted test datasets, and the use of non‑generative AI help reduce risks around bias, discrimination, and explainability. Governance measures, including RAI engagement and accuracy testing, have been proactively undertaken, with mitigations in place to support reliability and fairness.
A multidisciplinary team is in place, supported by training and additional agreements in progress. While some platform technologies are new to HMCTS, MoJ will provide short‑ to medium‑term support. A RACI model is being developed, and future BAU resourcing will be required post‑launch. Microsoft Azure, the strategic cloud provider for HMCTS, acts as the primary external supplier, ensuring the platform operates within an approved and secure environment.
5.2 - Risks and mitigations
These are mainly covered by the Responsible AI Assessment. All of the risk and mitigations are understood and have been addressed.
The areas used by the RAI process are:
- Lawful: The use of AI will comply with all applicable laws, standards and regulations Proportionate: AI systems will not make decisions where it is essential to have human oversight to ensure we administer justice fairly and uphold judicial independence
- Fair: AI systems will treat all people fairly and impartially, and will not discriminate unfairly against individuals or organisations
- Reliable: AI systems must function consistently in a robust, secure and safe way, be grounded in our own data, and comply with organisational technology standards
- Secure: AI systems will protect privacy by ensuring data quality, availability and security in accordance with privacy laws
- Explainable: People will have access to, and be able to understand, information on how AI systems operate, enabling individuals affected by AI systems to understand their outputs
- Transparent: We will provide public access to relevant information about AI systems in use across the courts and tribunals, while protecting valid security and intellectual property concerns
- Accountable: The use of AI will be properly governed, with clear lines of accountability identified to ensure effective human oversight of AI systems throughout their lifecycle
- Contestable: The use of AI in the courts and tribunals should be open to challenge, with proportionate and accessible processes in place for people to contest the outputs of an AI system
The summary of the ORIGINAL assessment findings were. All were LOW risk except Fair scoring MEDIUM.
The Medium Risk score is a result of some areas having not been considered in enough depth at this stage, as evidenced by the assessments yet to be completed. It is not because the proposal here appears inherently risky, and once the additional mitigations have been completed the overall risk score should reduce to LOW RISK. This risk score is for phases 1 and 2 of the project - AI transcription of judicial audio for judgments. The following actions should be taken to mitigate risk as part of pilot delivery (not exhaustive):
- Human Rights Impact Assessment and risk assessment against the government’s automated decision-making framework should be completed
- Document a process for ongoing testing/validation of outputs and for reviewing and changing AI models
- Further work to be undertaken to investigate how the data retention policy applied to the use case may be affected by the Record Retention and Disposition Schedules
- Judicial training provided should include clear information on how the AI works
- Seek direction on how to implement components which do not interact directly with the public
- Consider RACI for roles and responsibilities and proposed SRO for the final component needs to be agreed