Backchannel Berlin (with Inworld)

A research afternoon on frontier speech, audio and language models, hosted by the Inworld team. End-to-end speech-to-speech models can now listen while they talk, handle interruptions and hold a consistent voice; getting them to frontier quality is still open research, and running them in consumer apps in every language inside a few hundred milliseconds is a second problem. The published question list is the agenda: what a full-duplex model should predict and at what granularity, where cascaded STT-LLM-TTS breaks down against end-to-end, how to keep an LLM's reasoning once you train it to listen and speak, how far synthetic dialogue goes, and which of batching, caching, quantization and speculative decoding actually matters under latency. Around 56 people registered. Free; exact location on sign-up. Why go: a genuine research discussion at this depth on speech models rarely happens outside a lab. Best for speech and audio ML researchers and inference engineers.