At Imagicle, we believe Artificial Intelligence (AI) can be leveraged to power better, more inclusive customer conversations for all. We also recognize that by applying this technology, we have a responsibility to mitigate potential harm. That is why Imagicle develops its AI-powered features in accordance with its responsible AI principles [the “Framework”], which are based on the principles of Transparency, Fairness, Accountability, Privacy, Security, and Reliability. Imagicle translates these principles into product development requirements that form part of the product development lifecycle, alongside our Security by Design and Privacy by Design processes.
Accordingly, Imagicle Voice Analytics features that leverage AI are built with transparency, fairness, accountability, privacy, security, and reliability at their core. Each feature powered by AI undergoes an internal AI impact review — an assessment of how the technical underpinnings of the functionality measure against these principles.
Imagicle Voice Analytics AI features are built with these principles at the center of how Imagicle delivers the AI-powered technology. This Technical Note describes more information about Imagicle Voice Analytics AI features, and how Imagicle responsibly leverages AI to deliver the functionality.
Feature Overview
Speech-to-Text Transcription
The Speech-to-Text Transcription feature converts recorded voice conversations into written, searchable transcripts using AI speech recognition. Transcription is available across multiple languages, allowing supervisors and quality teams to read, search, export, and share conversation content rather than listening to recordings in full. The resulting transcripts form the basis for downstream analysis such as sentiment, summarization, and topic detection.
AI Sentiment Analysis
The AI Sentiment Analysis feature applies natural language processing to transcribed conversations to detect positive, negative, and neutral sentiment. Sentiment can be assigned to an entire conversation as well as to individual speakers and individual sentences, giving supervisors deeper insight into customer satisfaction, agent performance, and overall conversation quality. Results are presented in dashboards and can be filtered by time period and user.
AI-Powered Summarization
The AI-Powered Summarization feature generates concise summaries of recorded calls. The feature automatically highlights key points, action items, and outcomes, reducing the manual effort required to review conversations and helping teams act on the content of a call more quickly. Users can review each summary before relying on it for follow-up activities.
Topic Detection
The Topic Detection feature analyzes transcribed conversations to identify and highlight the most relevant topics discussed. By surfacing recurring themes — such as common complaints, service gaps, or frequently raised requests — the feature helps supervisors understand what matters across a large volume of conversations at a glance, without reviewing each recording individually.
Keyword and Sentiment Alarms
The Keyword and Sentiment Alarms feature allows supervisors to define conditions — such as specific keywords or sentiment thresholds — that trigger notifications when met within a conversation. This enables teams to act in near real time on interactions that require attention, supporting coaching, escalation, and quality-assurance workflows.
Model Overview
Introduction
Imagicle Voice Analytics AI features are powered by AI technologies that combine speech recognition with natural language processing and large language models (LLM). Certain functionality relies on third-party AI services: speech-to-text transcription is provided by Soniox STT, and the AI features that operate on transcripts are accessed through OpenRouter, an AI routing service that proxies requests to a range of underlying third-party models and providers. These services are operated within a cloud environment configured by Imagicle. For more information about the underlying third-party services, see the Soniox and OpenRouter transparency documentation referenced below.
Model Architecture
Imagicle Voice Analytics leverages AI speech-to-text models to transcribe recorded conversations and natural-language and large-language-model technologies to perform sentiment analysis, summarization, topic detection, and PII redaction on the resulting transcripts. Speech-to-text transcription is provided by Soniox STT. The large-language-model features — including summarization, topic detection, and PII redaction — are accessed through OpenRouter, an AI routing service that proxies requests to a range of underlying third-party models and providers. Imagicle does not develop the underlying foundation models; it configures and orchestrates them to deliver the Voice Analytics functionality.
Usage Guidelines
Imagicle Voice Analytics AI features operate on conversations captured by Imagicle Call Recording and are available as a cloud-native add-on to that service. The features are intended for use by authorized supervisors, quality teams, and administrators for purposes such as quality assurance, agent coaching, and compliance, and are not intended to make automated decisions about individuals without human review.
Model Inputs and Outputs
Speech-to-Text Transcription
A typical input is the audio of a recorded conversation. A typical output is a written, time-aligned transcript of that conversation in the detected language, which users can search, export, and share.
AI Sentiment Analysis
A typical input is the transcript of a recorded conversation. A typical output is a sentiment classification — positive, negative, or neutral — assigned at the level of the whole conversation, individual speakers, and individual sentences, together with aggregated trends shown in the dashboard.
AI-Powered Summarization
A typical input is the transcript of a recorded conversation. A typical output is a concise summary highlighting key points, action items, and outcomes. Users are encouraged to review the summary for accuracy before relying on it.
Topic Detection
A typical input is one or more conversation transcripts. A typical output is a set of detected topics or themes (for example, “billing issue,” “cancellation request,” or “technical support”) used to highlight and group conversations for supervisors.
Keyword and Sentiment Alarms
A typical input is a conversation transcript together with supervisor-defined conditions, such as a keyword or a sentiment threshold. A typical output is a notification raised when a conversation meets one or more of the defined conditions.
Data Sources for Training and Evaluation
Imagicle does not use customer content to fine-tune or train the foundation models used to deliver Voice Analytics AI features. Customer conversations are processed solely to generate the requested output (transcription, sentiment, summarization, topic detection, or alarms) for the customer.
Where Voice Analytics relies on a third-party AI service, the provider — Soniox STT for transcription — represents that it does not use customer data to train or improve its foundation models, and does not retain customer data passed to the service beyond what is necessary to provide the functionality. For more information about Soniox STT’s data, privacy, and security practices, see the Soniox transparency documentation referenced below.
The underlying foundation models are trained by their respective providers on large datasets compiled from a wide range of publicly available written and spoken material. This training data encompasses a diverse array of topics, languages, and styles to help the models develop a broad understanding of language and its contexts. For further information, see the provider documentation referenced below.
Model Evaluation and Performance
Imagicle periodically assesses the AI features to maintain and improve both the performance and accuracy of the generated output. Humans participate in the review, testing, and quality assurance associated with the deployment of the underlying models. Transcription, sentiment, summarization, and topic outputs are probabilistic and may contain errors; accordingly, the features are designed to support — not replace — human review.
Safety and Ethical Considerations
AI models may produce inaccurate, incomplete, or unintended output if such patterns exist in the input or the model. To mitigate this, the third-party AI services used by Voice Analytics include content-filtering and safety controls, and Imagicle constrains how the models are used to the specific, narrow tasks of transcription, sentiment, summarization, topic detection, and alarms.
In addition to these technical mitigations, Imagicle has taken further steps to support safe and ethical use. Imagicle’s third-party AI vendors undergo vendor reviews that include security and safety assessments. Output generated by Voice Analytics features is visible only to authorized users within the customer’s organization, unless shared by those users voluntarily. Please contact Imagicle if you have concerns or feedback about your experience with Voice Analytics.
Fairness
To understand how the foundation models underlying Voice Analytics are trained and evaluated for fairness, please reference the transparency documentation of the relevant AI service providers (Soniox STT and OpenRouter), referenced below. Imagicle supports transcription and analysis across multiple languages and reviews feature output to help identify and reduce systematic disparities in performance.
Privacy and Security
Imagicle Voice Analytics processes recorded customer conversations, which may contain personal data. Customers remain the controllers of the conversations they choose to record and analyze and are responsible for obtaining any required consents and for configuring recording and retention in line with applicable law and regulation (for example, MiFID II or PCI-DSS where relevant).
Recorded conversations are encrypted at rest, and access to recordings and analytics is restricted to authorized users. Imagicle addresses the security of customer data in its security and data-processing documentation, available on request. Where third-party AI services are used, the providers (Soniox STT and OpenRouter) represent that they do not access, monitor, or store customer data beyond what is necessary to provide the service.
Data Location
Imagicle Voice Analytics is delivered from regional data centers, and the location in which customer data is stored is determined by the region selected when the customer is onboarded. Customers onboarded in the European region — including customers located in the European Union and in the Middle East — are hosted in a data center located in the European Union. Customers onboarded in the United States are hosted in a data center located in the United States.
Recorded conversations, transcripts, and analytics output are stored in the data center of the customer's assigned region. Where Voice Analytics relies on third-party AI services, as described in the Model Overview section, processing by those services is governed by the providers' data-processing terms. For more information about data locations and sub-processors, see Imagicle's security and data-processing documentation, available on request.
Updates and Maintenance
Ongoing Imagicle product updates and release notes describe product changes and improvements to the user experience. Imagicle may also use customer email notifications and its support and community resources to inform customers of changes to Voice Analytics AI features.