Convo AI World

What does it take to translate a conversation as it happens? In this special montage episode of the Convo AI World Podcast, we bring together selected conversations with Palabra's Artem Kukharenko, Soniox's Klemen Simonic, Rime's Lily Clifford, Akool's Jeff Lu, and Professor Mérouane Debbah to explore the challenges of real-time translation. The episode begins with the human promise of speaking across language barriers, then examines the decisions a translation system must make before a sentence is complete. Klemen explains the tension between responding quickly and waiting for enough information to preserve meaning. Palabra's perspective adds the reality of mixed-language conversations, where people can switch between languages without announcing the switch. Lily and Mérouane broaden the discussion to dialects, regional data, and cultural context. Their insights show why supporting a language involves more than adding its name to a list. Jeff takes the conversation beyond words, exploring how translated video can preserve a speaker's voice and facial expression. From everyday multilingual interactions to enterprise requirements around privacy and customization, the masterclass connects technical choices to the experience of the people using them. It closes with a vision of translation becoming a seamless part of conversation and the work still needed to get there. Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world

In this episode of the Convo AI World Podcast, Hermes Frangoudis speaks with Neal Patel, co-founder of ScaleDown, about choosing models for the specific jobs inside an AI application. Neal explains how his work forecasting data center demand led him to focus on token efficiency. At ScaleDown, that means building small language models for tasks such as classification, extraction, and summarization. He describes how the team combines existing model components with task-specific neural heads, and where these specialists can work alongside a larger model that handles planning and conversation. For voice AI builders, the discussion covers routing, parallel subagents, conversation history, and post-call analysis. Neal explains the difference between query-aware compression and a generic summary, and why preserving the information needed for an answer matters more than simply shrinking a prompt. He also explores using classifiers for evaluation, where consistency can matter as much as speed. The conversation closes with ScaleDown's integration tools and roadmap, plus Neal's view of how model selection could become more automated. Throughout, he makes a useful distinction: specialized models may reduce variability on a defined task, but they are not fully deterministic. Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world

In this episode of the Convo AI World Podcast, Hermes Frangoudis sits down with Dustin Allen, co-founder of Trinitite, to explore why evaluating an agent is only part of making it reliable in production. Dustin traces Trinitite's origins to a productivity app for accountants and finance teams, where a damaged spreadsheet exposed the need to replay actions, restore data, and understand what went wrong. He explains why his team now treats AI delivery, governance, and security as connected parts of the same system, and how Trinitite uses deterministic inference to make its own assessments more repeatable and failures easier to investigate. The conversation gets practical about voice AI: where to place controls in a cascading pipeline, how asynchronous monitoring can preserve conversational flow, and why sensitive tool calls may need checks before an action proceeds. Dustin also discusses multimodal systems, developer tools, and the challenge of observing agents running locally on a device. For teams building voice agents or enterprise AI, the episode offers a grounded discussion of continuous monitoring, remediation, and the limits of guardrails, with an emphasis on learning from failures and keeping people involved in deciding which risks matter most. Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world

What does it take to give AI a face, and build a digital character people actually want to talk to? In this special montage episode of the Convo AI World Podcast, we bring together experts from Trulience, Akool, and AKA Virtual to unpack the technology, design choices, and product thinking behind the rapidly evolving world of AI avatars and digital humans. Jeff Lu of Akool explains why his team has focused on photorealistic people, real-time video generation, and running increasingly capable AI experiences closer to the edge. Jia Shen of AKA Virtual makes the case for a very different approach: characters do not always need to imitate humans. He explains why caricature, animation, and clearly defined character roles can sometimes create more believable experiences. Richard Bowdler of Trulience explores another part of the stack: giving conversational AI a face and body while keeping the underlying intelligence modular. That means combining avatars with LLMs, speech technologies, knowledge, and other components while allowing developers to choose the "brain" behind the character. Whether you're building conversational AI agents, virtual assistants, interactive entertainment, customer experiences, or the next generation of AI characters, this masterclass offers a practical look at what it takes to turn an AI model into a digital presence people can actually interact with. Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world

What does it take to evolve from live voice experiences to AI-powered game hosts? In this episode of the Convo AI World Podcast, Agora's John Moore sits down with Geremie Camara, co-founder of Live Play Mobile, to explore how conversational AI is reshaping live gaming and interactive entertainment. With more than 100,000 hours of live-hosted gameplay, Live Play Mobile has spent years learning what keeps players engaged. Now, the team is applying those lessons to AI avatars that can entertain, personalize experiences, and scale globally alongside human hosts. Geremie shares why real-time infrastructure matters, how AI and human hosts complement each other, what it takes to make AI feel natural, and the engineering challenges behind voice quality, latency, memory, and production-ready conversational AI. Whether you're building voice AI, AI agents, gaming experiences, or real-time applications, this conversation offers practical insights into where conversational AI is heading, and what it takes to build engaging AI experiences at scale. Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world

In this episode of the Convo AI World Podcast, Hermes Frangoudis sits down with Luka Chkhetiani from AssemblyAI to unpack what it takes to build reliable speech AI for real-world applications. Luka shares how AssemblyAI's mission evolved from making speech recognition accessible to developers into building robust models for real-time transcription, voice agents, conversational intelligence, and messy production environments. The conversation explores why latency alone is the wrong obsession, how turn-to-turn responsiveness changes the user experience, why reliability is now one of the hardest problems in speech AI, and what developers often misunderstand when bringing voice agents into production. Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world

In this episode, host Hermes Frangoudis interviews Faraz Siddiqi, co-founder and CTO of Bluejay, about testing and monitoring voice AI agents. Faraz explains how Bluejay uses "digital humans" synthetic customers with accents, background noise, and even emotions to run hundreds of parallel simulations. What used to take 10–12 hours of manual calling to test 200 menu items now takes just five minutes, letting you grab coffee while Bluejay finds failures before real users hit them. He shares why latency over 1.5 seconds breaks trust, how a transcription error turned "Dr. Pham" into "Dr. Fan," and why founders should listen to every call manually before automating. A must-watch for anyone building production-grade voice agents. Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world

In this special montage episode of the Convo AI World Podcast, host Hermes Frangoudis brings together leading researchers and founders from across the speech‑to‑text space to unpack the voice pipeline from the ground up. The conversation covers how cascading architectures stack up against real‑time speech‑to‑speech systems, why Voice Activity Detection acts as the critical traffic controller, and how enterprises can eliminate model pathologies like hallucinations and omissions through modular “Lego block” integration. Guests from Deepgram, Agora, Soniox, and Rime share hard‑won lessons on achieving near‑native accuracy across 60 languages with self‑supervised learning, taming unpredictable pronunciations in LLM‑driven agents, and why truly human‑like emotional understanding is still around the corner. The episode confronts the persistent myth that speech recognition is a solved problem, spotlighting the long tail of accents, rare words, and noisy real‑world conditions that still break most systems, and makes the case that for regulated, high‑stakes industries the auditable text backbone of cascading pipelines remains essential even as speech‑to‑speech models race toward a more natural future. Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world

In this episode, host Rishi Ahluwalia interviews Ankur Edkie, CEO of Murf AI, about the hidden complexities of scaling enterprise voice AI. Ankur discusses the journey from pre-ChatGPT text-to-speech to building hyper-efficient, human-like voice systems. He unpacks why voice demos feel magical, but production deployments often fail, highlighting the critical gap between lab settings and the variability of real-world acoustics and devices. The conversation covers the challenges of the cascading stack (ASR, LLM, TTS), the importance of turn-taking and latency consistency over raw speed, and Murf’s "compute acquisition" approach with the Falcon architecture that drives down costs and enables global data residency. Ankur concludes with a hard lesson for builders: enterprises are buying holistic outcomes and trust, not just isolated API calls. Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world

In this episode, host Derek interviews Zeyi, Founder & CEO of WaveSpeedAI, about simplifying AI media generation. Zeyi explains how WaveSpeed provides a unified, high-concurrency API that lets developers easily switch between models like Flux and Wan without managing complex GPU infrastructure. He highlights unique offerings like the 10-minute lip-sync tool InfiniteTalk and video extending, discusses regional preferences in AI-generated faces, and shares cost-saving strategies like low-res generation with upscaling. The episode wraps with Zeyi's advice for AI builders: leverage AI coding tools, structure projects wisely, and use API platforms instead of optimizing models from scratch. Check out video episodes and subscribe to the Convo AI Newsletter at podcast.convoai.world