Speech Analytics: Transforming Voice Data into Business Action

© 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Speech Analytics: Transforming Voice Data into Business Action November 2025 CRYSTAL RTTH Fall School -BILBAO 2025 Dayana Ribas, Ph.D. Lead Scientist SmartVoice © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Introduction Speech analytics Strategic aplication into business OUTLINE Intro, definition, antecedents, methods Use cases, results, proyections 1BTS and SmartVoice 2 3 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 3 BTS OVERVIEW 3 BTS Headquartered in Miami, with operations center in Zaragoza, Spain Ownership BTS Group 100%, +160 employees Solutions Connectivity (Voice, A2P Messaging, Cloud Numbers, Programmable Communications, eSIM), SmartVoice, Protection, Managed Communications, Omnichannel, Identity, and Analytics Platforms Proprietary S1 Platform for Voice, Messaging, Protection, Managed Communications, Identity, and Analytics Segment Global solutions for Telcos, MNOs, CpaaS, and Hyperscalers Website www.bts.io Value Propositio n With over 30 years of experience, BTS is a leading connectivity technology provider and partner for CommTech players worldwide. Telegeography Ranking Carriers, 2023 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 4 SMARTVOICE From legacy service to an intelligent layer: SmartVoice SmartVoice is BTS’s advanced audio and voice processing unit, powered by AI. It delivers cutting-edge solutions for speech analytics, enhancement, transcription, and voice biometrics, turning raw audio into actionable insights. Designed to improve communication quality, customer experience, and business intelligence, SmartVoice enables organizations to truly understand and optimize every conversation. AI-first, network-native: •Tested at scale in contact centers for specific use cases. •Moved into the network for in-path analysis and real-time decisions •Works over current connections; no new circuits or deployments •Unlocks a portfolio of AI-based services for operators, enterprises, and platforms over a single connection. Capabilities across the call path: Analyze (emotion, conversational flow, quality, speech profiling), Generate (text-to-speech), Act (on-network voice agents), Verify (voice biometrics with anti-spoofing), Enhance (real-time speech denoising). PCT/EP2025/069103 “A SYSTEM AND METHOD FOR ENHANCING A CALL CHARACTERIZATION OF VOICE CALLS” 4 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 20252025 55 2025 SmartVoice -Funtionalities CONTACT CENTER APPLICATIONS 1) Audio and Speech Analytics 2) Speech Enhancement 3) Voice Biometrics 4) Speech to Text BTS SMARTVOICE | FUNTIONALITIES VOICE MANAGEMENT APPLICATIONS 1) Voice Traffic Profiling 2) Real Time Speech Enhancement 3) Deepfake detection 4) AI-Driven Phone Call Assistants 5) Customer Service Assistance Tools that support functionalities: Audio diarization: Who spoke when? | Voice Activity Detection (VAD) | Antispoofing system: Deepfake | Language Identification | Music segmentation | Speech Repetition Detection © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 20252025 Language is just words? 016 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 In English, time is usually spoken of in horizontal terms. (“What lies ahead,” “leave the past behind”) ENGLISH MANDARÍN In Mandarin,by contrast, a vertical framework is used (“up = before,down = after”) Boroditsky, L. (2001). Does language shape thought? Mandarin and English speakers’ conceptions of time. Cognitive Psychology, 43(1), 1–22. LANGUAGE METAPHOR OF TIME Spanish Horizontal (forward = future, backward = past) English Horizontal (forward = future, backward = past) Mandarin Horizontal + Vertical (up = future, back = past) Arabic / Hebrew Horizontal but right, left in writing This even affects how people arrange event cards chronologically:English speakers do it in astraight line, while Chinese speakers do it from top to bottom. Semantics organizes the way we think about business problems. The language shape the time concept 7 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. Winawer, J., et al. (2007). Russian blues reveal effects of language on color discrimination.PNAS, 104(19), 7780–7785. 8 The language shape the color indentification LANGUAGE SHAPES THE PERCEPTION OF COLOR Participants 13 Russian speakers (living in the U.S. but Russiandominant), 13 English speakers monolinguals All had normal color vision (tested with standard Ishihara color plates). Participants were matched for age and education. Results If language influences perception, of blue Russian speakers should be faster at distinguishing two shades hat cross the goluboy/siniy boundary than two shades within the same category while English speakers should not show this advantage. Russian speakers responded significantly faster, in the right visual field. The effect disappeared under verbal interference Confirming a language-based mechanism © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Two groups of persons Same text THE ONLY DIFFERENCE: ONE KEY WORD VIRUS Provides proactive and effective solutions (education, community programs) Thibodeau & Boroditsky, L. (2011). Metaphors We Think With: The Role of Metaphor in Reasoning. When asked, most denied that a single word influenced their decition… BUT IT HAD! BEAST Leads to punitive reactions and harsher punishments. THE LANGUAGE MODIFIES EMOTIONS 9 The words employed shape interpretation and emotionality of the listener © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 It analyzes the textual content of conversations: words, key phrases, topics, requests, or complaints. It focuses on the explicit intention of the customer. Allows detecting topics such as: Reason For Frequently Asked Service Contact Questions Barriers Terms Associated with Churn or Purchase Example:“I want to cancel my service” →Detects churn intent. WHAT SPEECH ANALYTICS CONSISTS OF 1 16 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 The goal is to identify the customer’s emotion, which allows detecting: Using voice signals, it is also possible to analyze tone of voice, speed, volume, pauses, and intonation. Is the conversation positive, negative, or neutral? Is the user frustrated, satisfied, or confused? Frustration, anger, calm, or enthusiasm Positive, negative, or neutral conversation Confidence or doubt Example:A phrase like “Yes, that’s fine” can sound resigned, ironic, or sincere depending on the tone. 2 WHAT SPEECH ANALYTICS CONSISTS OF 17 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Main KPI Performance Productivity Satisfaction 123 It analyzes the average response time, the first contact resolution rate, the frequency of repeated questions, and the problem escalation rate,whether to another agent or to a human if it was abot. For the quality evaluation of the conversation, we assess: •The fluency of the dialogue •The clarity of the responses •The coherence and naturalness Example:Average response time of the assistant in replying & % of inquiries resolved without human intervention. 3 WHAT SPEECH ANALYTICS CONSISTS OF 18 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Modelo ¿Para qué sirve? Intent Classification Models Understand what the user wants Sentiments and Emotion Analysis Models Detect how the user feels Entity extractions Models Identify key information Clustering and topic modeling Group similar conversations and topics Summarization model Generate short clear summaries for dialogues Anomaly detection model Used to spot unusual interactionss Conversational quality socoring model Evaluate the fluency and conversations AI MODELS –SPEECH ANALYTICS Types of Models Used 19 Conversational analytics A very illustrative topic to talk about AI in this moment. © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. Transcription A common starting point, as it allows understanding what was said, facilitating searches, categorization, and content analysis.Tone of Voice Measures Reveals Frequency of vocal cord vibration Emotional satate: anger calm, doubt Volume / Intensity Audio amplitude Nervousness, aggresiveness, low energy Pauses / Silences Gaps in speech Waiting time, hesitation, technical cuts Overlaps / Interruptions When both speakers talk at the same time Lack of listening, poor coordination Saturations / Noise Level of clarity and cleanliness of the comunication channel Technical quality of services Beyond the text of the transcription, VOICE –CHALLENGES AND OPPORTUNITIES 20 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 20252025 Application into business 0321 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Contact Centers Voice-base Sales Healthcare Banking Analysis of the salesperson´stone and the customer´slevel of engagement. Detection of emotional or cognitive deterioration in patients. Voice identification and intelligent audio monitoring. Realtime detection of dissatisfied or upset customers. SPEECH ANALYTICS –USE CASES 22 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Privacy and security in conversational analytics are critical concerns, especially because they involve the processing of highly sensitive data, such as: Personally Identifiable Information (PII), emotional and psychological content, financial or health data, opinions, and detectable biases in conversations. Main aplicable regulations Other standars and frameworks Anonymization Informed Consent Retention limitation Data-at-rest/ In-transition encryption Granular access Data protecton Impact assessment (DPIA) Differential privacy models Explainability Stay Secure, Talk Smart •Exposure of sensitive data •Re-identification •Improper data retention or use •Algorithmic biases and automated decisions RISKS ASSOCIATED WITH CONVERSATIONAL ANALYSIS SECURITY AND PROVACY TECHNIQUES Tom Backstrom (2023) “Privacy in speech technology”, Procedings of the IEEE. SPEECH ANALYTICS –SECURITY & PRIVACY 23 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 • Offers regional storage options. • Export only under configured security policies. • Integrates with Data Loss Prevention (DLP) API to automatically detect PII. Examples • Automatic redaction of PII data (optional). • Supports compliance with HIPAA,GDPR, and SOC 2. • Encryption of conversations at rest with customer-managed keys (KMS). CCAI Architecture 1 2 3 45 6 Design “privacyby-design” from the start of the system. Apply automatic filtering of sensitive data before training models. Implement rolebased access control system. Presiodically Audit models to detext biases or data leaks. Include explicit consent options in conversational interfaces, (opt-in) Inform the user if their conversation Will be analyzed (e.g., workplace, health eviroments) Recommendations of good practices SPEECH ANALYTICS –SECURITY & PRIVACY 24 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 20252025 25 1. Use case and project definition 2. Models and APIs 3. Orchestration 4. Infrastructure 5. Client-side tools The industrialization of a system is composed of © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Transcription of the calls Audio of the call by speaker Alarms detection of alarms in the call INTO THE SOUND | CLIENTSIDE TOOLS | AUDIO & SPEECH ANALYTICS © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Calls IPA SATISFACTION INDEX SPEECH LEVEL NOISE FLOOR EXCESSIVE NOISE SPEAKER TURN DURATION MAX RESPONSE TIME SILENCE INTONATION (MONOTONOUS) SPEAKING RATE (SLOW) Call Center 1 846 (7,7%) 52,1% 94,2% 2,8% 0,1% 2,0% 2,5% 1,2% 1,7% 3,6% 2,1% Call Center 2 8.028 (73,5%) 58,5% 92,5% 1,1% 5,0% 5,8% 5,0% 0,9% 5,2% 4,2% 3,3% Call Center 3 985 (9,0%) 52,1% 93,3% 4,7%0,0% 0,2% 0,8% 0,6% 3,4% 3,6% 1,9% Call Center 4 1.066 (9,8%) 56,3% 93,1% 3,5% 0,0% 0,3% 1,1% 0,9% 4,2% 2,4% 2,2% TOTAL 10.925 57,2% 92,8% 1,7% 3,7% 4,5% 4,0% 0,9% 4,6% 3,9% 3,0% COMPARATIVE CALL CENTERS Channel Agents AUDIO QUALITY CALLFLOW PARALINGUISTIC •Call Center 2 has the lowest Satisfaction Index, being also the one with the greatest number of alarms that impact this indicator. •Call Center 3 presents a high number of speech level alarms compared to the rest 1 ene –22 feb Analysis Period 33 REPORT –CUSTOMER SATISFACTION (EXAMPLE) INTO THE SOUND | CLIENT-SIDE TOOLS | AUDIO & SPEECH ANALYTICS © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Results 34 HIGH SENSITIVITY TO QUALITY ISSUES: Over 52.9% of analyzed calls triggered qualityrelated alarms, with audio issues accounting for 34.8% of these alarms. IMPROVED KPI TRACKING: Metrics such as the Quality Index (QI) and Satisfaction Index (YES) revealed variations in performance across different contact centers, enabling targeted interventions. For example, Contact Center 2 was identified as requiring significant improvement. ACTIONABLE INSIGHTS: ITS identified specific recorder and switch issues, directly aiding technical teams in resolving longstanding operational problems. ENHANCED CUSTOMER SATISFACTION: The analysis linked audio quality improvements to higher satisfaction rates, validating ITS’s impact on customer experience. 34 © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. INTO THE SOUND | RESULTS | AUDIO & SPEECH ANALYTICS The ITS Toolkit has demonstrated measurable success during its implementation with Telefónica's contact centers. © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. Phase Academic Focus Industrial Focus Key Output Research Discovery & publication Validation of novelty Scientific proof Proof of Concept Prototype development Technical feasibility Lab-scale demo Technology Transfer Project & partnership Market alignment License or spin-off Industrializa tion Scale-up engineering Commercial viability Product or process Protection Patent strategy Legal exclusivity Asset “What is scientifically interesting?” to “What creates measurable value in the market?” The patent acts as a bridge and a shield, protecting the original innovation while enabling partnerships, funding, and large-scale deployment. INTO THE SOUND | INDUSTRIALIZATION OF A METHODOGY INDUSTRIALIZING AN ACADEMIC METHODOLOGY REQUIRES SHIFTING FROM 35 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 From Idea to Patent: Industrialization of a Methodology INTO THE SOUND | PATENT PROCESS CONCEPTION & PATENTABILITY ASSESSMENT Originates from R&D teams (scientists, engineers, designers). Evaluate 3 patentability criteria: Novelty:Must be new, not previously disclosed. Inventive Step: Not obvious to experts in the field. Industrial Applicability:Must have practical use. Conduct prior art search in patent databases and literature. If innovative and strategically valuable → proceed with patent filing. DRAFTING THE PATENT SPECIFICATION Title & Abstract:Short summary of the invention. Background:Define the problem and existing limitations. Detailed Description:Technical explanation (inputs–outputs, level of detail trade-off). Figures: Visuals or schematics of the invention. Claims:Define the legal protection scope (drafted by patent attorneys). Language must be broad enough to protect, yet precise enough for legal validity. PUBLICATION STRATEGY & REVIEW PROCESS. Patent remains confidential for 18 months, then published automatically. Examination phase: •Examiner checks novelty and inventive step. •Applicant responds to official actions until approval. Strategic considerations: •Avoid public disclosure before filing. •Build a patent portfolio around key technologies. DURATION & COSTS Protection period:20 years from filing date. Maintenance:Regular renewal fees required. Typical costs: •National patent:5,000 – 15,000 eur •U.S. / Europe:10,000 – 30,000 eur •International (PCT): 30,000 –100,000+ eur 36 © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Speech Profiling of Telephone Calls 3737 © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Studies revealed and commented back in July Categorization of phone calls Objective: Identify conversational vs. non-conversational calls Reports identify conversational ratio INTO THE SOUND | PIVOTING & EVOLUTION © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 38 Models balancing efficiency and effectiveness Cost of storing new data at these volumes Adaptation to call peaks Real-time processing capacity (delays <50ms) Volume and Storage Security and Privacy Encryption of sensitive data Pseudo-anonymization of personal data Backup and recovery copies Facing Challenges: INTO THE SOUND | INNOVATION FRAMEWORK © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Tools and frameworks for analysis •Chang et al. (2020) –ConvoKit: A Toolkit for the Analysis of Conversations Framework open-source para representar, explorar y analizar grandes colecciones de datos conversacionales Reddit+15arXiv+15MDPI+15ScienceDirect. Survey studies (≥ 2021) •Gao et al. (2021) –Advances and Challenges in Conversational Recommender Systems: A Survey Revisión detallada del estado del arte en sistemas de recomendación conversacional: estrategias de diálogo, interrogación del usuario, explotación-exploración y evaluación arXiv. •Soudani et al. (2024) –A Survey on Recent Advances in Conversational Data Generation Explora métodos recientes para generación de datos sintéticos de diálogos, datos multi-turn, filtrado de calidad y desafíos metodológicos arXiv+1arXiv+1. •Zafar et al. (2023) –Building Trust in Conversational AI: A Review and Solution Architecture Arquitectura que combina LLMs con grafos de conocimiento para sistemas conversacionales explicables y con enfoque en privacidad y confianza arXiv. •Singh et al. (2024) –revisión integral sobre chatbots IA-powered: aplicaciones, desafíos técnicos, bias, ética y direcciones futuras; cubre desde ELIZA hasta ChatGPT y modelos recientes ScienceDirect. Industry ans market reports •DMG Consulting (2024) –Conversation Analytics for the Digital Era (Julio 2024) Análisis del mercado actual de analítica conversacional orientada a contact centers: capacidades de IA/GenAI, resumen automatizado, detección de emociones, AQM, RTG… dmgconsult.com. •StartUps Insights / Research & Markets (2025) –Market Survey & Forecasts (2024-2029) Información del crecimiento del mercado, tendencias en modelos híbridos generativos/discriminativos y evolución tecnológica globenewswire.com+1startusinsights.com+1. First approaches: Weizenbaum (1966 ELIZA), Radford et al. (GPT-1, 2018) Academic surveys: Gao et al. (2021), Soudani et al. (2024), Zafar et al. (2023), Chen et al. (2018) Practical toolboxs: ConvoKit (2020) Privacy and Security: Tom Backstrom (2023) “Privacy in speech technology”, Procedings of the IEEE. References © 2024 BTS. All rights reserved. Contains Confidential and Proprietary Information of BTS. © 2025. BTS Inc. All rights reserved. Contains Confidential and Proprietary Information. 2025 Thank You [email protected]