Almost every voice product you have used is a committee. One model transcribes what you said. A second one thinks up a reply. A third turns that reply into sound. They are made by different companies, billed on different meters, and glued together with code that hopes the seams do not show. Kalpa Labs, a San Francisco lab that went through Y Combinator's Fall 2025 batch, thinks the committee is the problem.
Its plan is to replace the whole arrangement with a single generalist speech model - one system that can take dictation, speak, clone a voice, and reason about audio, all from the same set of weights. The reference point the founders keep coming back to is text. Before GPT-3, natural language processing was a drawer full of narrow tools. After it, one model with good instructions could do most of the drawer's work and learn new tricks on the fly. Kalpa Labs is betting speech is standing where text stood a few years ago.
01 / The ProblemA stack held together with tape
The speech AI market grew up specialized. There are excellent companies that only do text-to-speech, others that only transcribe, others that only clone a voice. Each is good at its job. Stacked together, they are less than the sum: every handoff adds latency, and context gathered in one stage rarely survives into the next. A voice agent that sounds calm when it greets you can go flat two sentences later because the model doing the talking never learned what the model doing the thinking meant.
Kalpa Labs frames this as fragmentation - too many models, too many vendors, and not enough actual listening. The company describes itself as an audio research lab working on models that understand nuance rather than just generate sound: catching a disfluency, a hesitation, a regional accent shift, the emotion sitting under the words.
Listening, in the company's telling, is the part everyone skipped. It is one thing to produce a clean, natural-sounding sentence. It is another to notice that the person on the other end trailed off, or switched registers, or said one thing while meaning another. Those are the signals a good human operator reads without thinking, and they are exactly what falls through the cracks between a transcription model and a text-to-speech model that never share a memory. Kalpa's argument is that comprehension and generation should not be two departments.
There is a business reason the fragmentation persists, too. Each specialist has an incentive to be excellent at its slice and indifferent to the rest, because the rest is someone else's product. That works for the vendors and quietly costs the builder, who inherits the integration, the latency budget, and the awkward moments when the pieces disagree. Kalpa Labs is aiming its whole design at that builder.
02 / The FoundersPlanet-scale ML meets nanosecond latency
The two people making this bet arrive from opposite ends of computing. Prashant Shishodia, the CEO, led full-stack machine learning on Google Assistant, where he trained and evaluated models and helped scale smaller Gemini variants to billions of queries a month. He has seen what voice looks like when a planet is talking to it at once. Gautam Jha, the CTO, spent his career in high-frequency trading, building systems where the unit of latency was the nanosecond.
It is an odd pairing until you notice what voice AI actually demands: models that are smart, and infrastructure that is unforgiving about speed. A conversation that lags feels broken no matter how clever the reply. Kalpa Labs sits at the intersection of those two habits of mind.
The trading background is not a footnote. Real-time voice lives or dies on the same discipline that markets do - shave a hundred milliseconds and the whole thing starts to feel alive instead of stilted. Someone who spent years hunting nanoseconds treats latency as a first-class problem rather than an afterthought to be optimized later. Pairing that instinct with a decade of watching how models behave at consumer scale is the kind of complementary bet investors tend to like.
It also explains the company's temperament. The name Kalpa comes from a Sanskrit word for an immense span of cosmic time, which is a big label for a two-person team - but the work so far is unglamorous and concrete: cheaper training, faster inference, published numbers. The ambition is long; the method is small and measurable.
Led full-stack ML on Google Assistant; trained and evaluated models and scaled Gemini variants to billions of queries a month.
Built nanosecond-latency systems in high-frequency trading before turning that speed obsession toward real-time voice.
03 / The MethodMaking audio as cheap to train as text
The most quietly interesting decision at Kalpa Labs is not about scale. It is about cost. Training on audio has long been expensive because of how sound gets turned into tokens a model can chew on - the residual vector-quantization approaches most labs use carry a heavy tax. Kalpa says it removed the bottlenecks in that step, which brought the price of training on speech down toward the price of training on text.
The proof point they cite is blunt: an 800-million-parameter speech model trained for under $1,000. Across the range, the company says it pretrained models from 800M to 4.8B parameters on roughly 2 million hours of mixed audio. Out of that came behaviors they did not hand-code - the models pick up regional accents, adjust prosody, and handle disfluencies from plain text, with no explicit tags telling them to.
04 / The ScorecardA number you can actually check
Claims in voice AI tend to be vague. Kalpa Labs published one that is not. In blind preference tests - listeners picking which clip sounds better without knowing the source - the company reports its model won 59.3% of the time against ElevenLabs' eleven-flash and 54.0% against eleven-turbo. ElevenLabs is the name most people reach for first in this category, which makes the comparison a deliberate one.
A falsifiable stat is a small act of confidence in a field thick with adjectives. It is also the kind of number a buyer can demand before signing anything. Blind preference is not the only thing that matters - latency, reliability, and cost per minute all decide real deployments - but it is the metric that maps most directly to whether a listener believes the voice. For a company whose whole thesis is that generalist beats specialist, beating the leading specialist at its own core task is the point that has to land first.
The honest caveat is that these are the company's own reported figures, drawn from its own testing. Independent evaluation is what turns a promising number into an accepted one, and that is still ahead. What the number signals for now is a willingness to be measured, which is not nothing in a category where most marketing pages prefer the word "lifelike" to any digit at all.
05 / The ProductsThree doors into one model
Rather than sit on a research demo, Kalpa Labs put three products into open beta, each a different way to reach the same underlying model. Studio is a scene-direction tool: you write dialogue, cast voices, set the mood, and generate multi-speaker scenes with emotion and consistent characters - closer to directing actors than typing into a text box. Realtime is a browser window for live, back-and-forth conversation with the model. And the API hands developers the same capabilities, with documentation the company says it wrote to be readable by both people and AI agents.
Shipping three surfaces before most seed startups ship one reads less like impatience and more like research. Each product is a way to learn what people actually want a speech model to do.
06 / The MarketBetting against specialization
Kalpa Labs is walking into a busy room. ElevenLabs, Cartesia, Deepgram, PlayHT and Hume all have real traction, and the biggest labs - OpenAI and Google among them - keep pushing their own speech features. Most of those players are specialists by design. Kalpa's contrarian claim is that specialization becomes a liability once a single base model gets good enough, the same way general-purpose language models eventually absorbed a shelf of narrow NLP tools.
Whether that thesis holds is genuinely unsettled. But the shape of the bet is clear, and the company is small enough to move on it. Backed by Y Combinator and, per reports, Nexus Venture Partners, Kalpa Labs is aiming at developers and companies building voice agents and conversational products - the people currently paying three vendors to do one job.
The business model follows the model itself. Kalpa Labs sells access: an API metered by usage, plus hosted tools like Studio and Realtime for teams that want to move without building their own front end. It is the familiar developer-platform shape, and it lets a two-person company reach customers far larger than itself. If the single-model approach genuinely cuts latency and cost while holding quality, the sales pitch is unusually easy to state - one integration, one bill, one voice all the way through the call.
There is a version of the next few years where speech quietly follows the path text already walked: a handful of general models absorbing a landscape of point solutions, with the survivors being whoever trained cheapest and steered best. Kalpa Labs is not guaranteed to be one of them. It is a small team, early, up against well-funded incumbents. What it has is a clear thesis, a couple of checkable numbers, and founders who have each seen a piece of this problem at a scale most people never will. In a field crowded with confident adjectives, a company arguing from arithmetic is at least asking the right question out loud.