Speech-to-text software Amberscript

Consumer and Market Authority

Speech-to-text software. The software uses AI to create a written record of interrogations and hearings. It does this by converting audio into written text.

Last change on 29th of July 2025, at 17:19 (CET) | Publication Standard 1.0

• Publication category: Other algorithms
• Impact assessment: Field not filled in.
• Status: In use

General information

Theme

Economy

Begin date

1-2023

Contact information

[e-mail]

Link to publication website

Bevoegdheden | ACM

Responsible use

Goal and impact

The algorithm is used by a hired party to convert audio (speech) into text. Citizens and companies participating in an interrogation or hearing encounter this algorithm as the audio recording of the interrogation or hearing they are participating in is converted to text by the algorithm. The generated text (the record) is always submitted to the parties for approval.

The impact on regulation, detection and monitoring is minimal because beyond converting sound to text, the algorithm does not perform any actions that affect citizens and businesses.

Considerations

The advantage of this algorithm is that interrogations and hearings no longer have to be typed out manually. This is reasonably justified as it ensures that recordings of interrogations and hearings are converted into text faster and cheaper.

Human intervention

Parties whose voice recording has been used are asked whether they agree with the elaboration. Thus, a record of an interrogation or hearing is always still submitted to parties.

Risk management

A report will be submitted to parties for agreement

Operations

Data

Audio recordings of interrogations and hearings.

Technical design

Speech-to-text (STT) software automatically converts spoken language into written text. It uses speech recognition technology for this purpose, which combines acoustic models and language models.

The acoustic model analyses the audio signal and converts it into feature representations - these are numerical summaries of the sound, such as pitch, loudness and frequency patterns, extracted in short time intervals (usually milliseconds). These features allow the system to recognise patterns in speech more efficiently than with raw audio alone. In traditional systems, these features are associated with phonemes, the smallest sound units in a language.

The language model then interprets the sequence of sounds by predicting the most likely word order based on grammar and context. It helps the system choose the right words, especially when sounds are unclear, similar or ambiguous (such as "to" vs 'two' vs 'too').

Modern speech-to-text systems often use end-to-end deep learning models, which handle the entire process - from raw audio to final text - using a single neural network. These models automatically learn both acoustic and linguistic patterns and typically achieve higher accuracy.

External provider

Amberscript Global B.V.