June 2026 Summaries

Voice AI technology is evolving from traditional pipeline systems to more advanced speech-to-speech models that can process and generate audio without relying on intermediate transcripts. While current pipeline systems are effective, they are limited by their inability to capture paralinguistic signals like tone and emotion, leading to potential misinterpretations in conversations. Speech-to-speech models promise to address these limitations by directly processing audio, allowing for more nuanced conversational interactions. However, these models present new challenges, such as the need for joint learning of perception and expression, the emergence of new failure modes, and the difficulty in training with conversational appropriateness due to the lack of natural supervision. The transition to speech-to-speech architecture requires rigorous evaluation of systems in realistic conditions to ensure they can maintain consistency, stability, and appropriateness over extended interactions. As the industry moves towards these advanced systems, the success of teams will depend on their ability to measure these capabilities accurately rather than relying on architectural promises alone.