Eight Years of Teaching Machines to Listen

In 2018, when we started Symbl.ai, investors found the pitch either obvious or absurd, with very little in between. Every organization runs on conversations. Sales calls, support calls, meetings, standups. And almost none of that is usable, because it evaporates the moment it happens. The most valuable data a company generates was, functionally, a waste product.

We bet that machines would eventually learn to understand human conversation, not just transcribe it, and that whoever built the infrastructure for that understanding would sit under an enormous amount of value.

Consider the timing though. The transformer paper was a year old. “Conversation intelligence” was not a category yet. Speech recognition had just gotten good enough that most people assumed transcription was the finish line, and chatbots were a running joke. Telling people that dialogue understanding would become one of the defining problems in AI earned polite nods and short meetings.

Eight years, a stack of patents, one production LLM, published research, and an acquisition later, that bet reads very differently. Voice agents answer real phone lines now. Every meeting tool ships an AI summary. Foundation labs compete on speech-to-speech models. Conversation turned out to be as central as we hoped, and harder than we knew.

This is what those eight years actually looked like, in three eras, and what each one taught me.

Era one: pipelines (2018–2022)

Before LLMs, understanding a conversation meant breaking the problem into tasks and building a model for each one. Take live audio, transcribe it in real time, then run purpose-built models over the stream. Detect the conversational structure, who is addressing whom, what is a question, what is an answer. Extract the topics actually being discussed. Catch action items as they form, so “I’ll send it Friday” gets flagged, assigned, and dated. Compress an hour of crosstalk into a summary a human would trust.

We built all of that against real-world audio, and several pieces of the work became patents, on conversational structure from speech in 2021, topics and action items in 2022, and summarizing conversational speech in 2023.

That era taught me disciplines I still use daily. Real-time systems keep you honest, because a meeting assistant that shows up four seconds late is not a meeting assistant, so every model had to earn its latency budget before it shipped. Errors compound, since ASR mistakes cascade into NLU mistakes, which made robustness to noisy input table stakes rather than an afterthought. And the biggest one, which took me years to fully appreciate, is that evaluation is the product. When you sell understanding, you have to define it, measure it, and defend the measurement to customers. We learned to build evals before features.

The limits of that era were structural though. Narrow models are controllable and cheap, but every new capability was its own research project, and we could feel the ceiling.

Era two: models (2022–2024)

Then large language models changed the equation.

Capabilities we had engineered as separate systems, like summarization, extraction, and question answering, became things a single model could do with a prompt. For a company built around such a pipeline, that raised every existential question at once. A lot of what we had built was becoming commodity, and a lot of what we had wanted to build was suddenly possible.

We responded in two ways, one with engineering and one with science.

The engineering answer was Nebula, our own LLM trained for conversation, shipped as a production API in 2023. I’ve written that story separately. The short version is that the durable value turned out to be the conversation-native data discipline, the evaluation suites, and a team that understood language models from the inside because it had built one.

The science answer came from a question that had nagged at me the whole time. Everyone assumed general models would simply absorb conversation like every other domain. Would they? We measured it, analyzing how model attention behaves on human dialogue versus web text, code, and math. The results backed up the practitioner’s intuition. Conversations demand measurably more complex, longer-range processing, and models specialize for code and math in ways they don’t for dialogue. The gap we had built a company on was real, and you could see it inside the models.

The lesson from that era, for me, is that platform shifts don’t kill fundamental problems, they kill implementations. Even as the pipeline evolved into something very different, our bet that conversation understanding is special, valuable, and unsolved got stronger.

Era three: scale (2025–)

In 2025, Symbl.ai was acquired by Invoca, where I now lead AI engineering.

Enterprise scale is a different sport, and I mean that as a compliment. When businesses route real revenue through your AI, when the phone call your model is handling is the sale, the tolerances change. Impressive most of the time is a demo. Production means predictable latency under load, quality that doesn’t drift silently between model versions, failure modes you understand and can bound, and evaluation rigorous enough to be treated like a contract. The startup years built conviction and capability. The enterprise years are where I’ve watched the last 10%, the reliability and trust and operational discipline, turn out to be most of the product.

What surprised me is that the founder’s thesis got sharper under enterprise constraints, not duller. Conversations are where revenue actually happens, and understanding them reliably at scale is worth more than any demo ever suggested.

What eight years taught me

Compressing ruthlessly:

1. Transcription was never the product. The words are what’s left over after a conversation. The product is what the words were doing, the commitments, decisions, objections, and trust. That’s as true for a 2026 voice agent as it was for a 2019 meeting summarizer.

2. Conviction is a time arbitrage. The market’s interest in conversation AI spiked in 2023 and again with voice agents now, but the problem was just as important in 2018. If a thesis only feels right when the market agrees with it, it was a trend all along.

3. Evaluation is the durable moat. Models commoditize every quarter. The ability to measure whether a system understood a conversation compounds over years, transfers across model generations, and is very hard to copy.

4. Conversation is a systems problem. Latency, interruption handling, memory, audio, state. The model is one component among many, and teams that treat voice AI as an LLM with a microphone tend to rediscover this in production.

5. The hard part was never the machines. Human conversation is messy in ways that are hard to overstate. It’s improvised, elliptical, emotional, and owned by everyone in the room at once. Teaching machines to meet that mess on its own terms is the actual job, and it is far from finished.

What comes next

Voice is becoming a default interface for AI, which means conversation stops being a data source and becomes the runtime. Agents are joining conversations as participants, first alongside humans and increasingly with each other, and dialogue itself will evolve in response. We’re already seeing early forms of human-agent discourse conventions. Speech-to-speech models are putting prosody, the channel transcription always destroyed, back into the loop. And memory across conversations will separate assistants that react from assistants that know you.

Every one of those developments raises the value of the same old unsolved problem. Does the machine actually understand what’s happening in the conversation? That question has survived three platform eras without losing any relevance. I’ve spent eight years on it, and I intend to spend the next eight on it too.

This site is where I’ll write down what I learn. The first interface humans ever built deserves machines that finally understand it.