Kalpa Labs: steerable conversational speech models in beta

Kalpa Labs is building generalist audio models that can listen, speak, follow instructions and learn new voice tasks from context. The current product is already more concrete than a research-only demo: a conversational beta in Studio, a live browser experience and an API for teams building on the model.

What it does

The current Kalpa homepage ↗ says teams can direct a multi-speaker scene in Studio, talk to a model live in Realtime and build through a REST API. Its stated research direction is an audio model that handles speech-to-text, text-to-speech, speech-in/speech-out reasoning, voice style and cross-modal tasks with instruction-following. The documentation site ↗ is linked as the build surface.

Kalpa’s launch report ↗ describes the problem as a fragmented speech stack: separate systems for transcription, synthesis, voice design, conversation and dubbing. It reports blind human-preference wins of 59.3% against ElevenLabs Eleven Flash and 54.0% against ElevenLabs Eleven Turbo for a beta TTS model. Those are company-published evaluation results, not an independent listening study or a guarantee for every voice and language.

Why I’d look closer

The advantage is steerability. A team can specify pacing, mood, pronunciation or delivery in context rather than picking from a fixed voice menu. The buyer could be a voice product, media team, game studio, customer-support application or developer experimenting with speech-native interfaces.

The tradeoff is reliability and permission. Generalist audio models need to preserve identity, follow a script, handle interruptions and avoid cloning a person without consent. The launch also says Kalpa is still investigating emergent behavior. Beta quality, latency, language coverage, rate limits and commercial voice rights matter more than a single preference number.

What I’d ask

Which API capabilities are stable enough for production? How are voice samples stored and deleted, and what consent record is required for cloning or transformation? Can developers control latency, turn-taking, pronunciation and output format? What happens when a prompt asks the model to imitate a living person or a copyrighted performance?

My editorial take

Shortlist Kalpa if your product needs speech that follows a scene or instruction, not just generic text-to-speech. Start in Studio or a low-stakes prototype, measure the voices and languages your users actually need, and keep identity-sensitive use behind explicit consent and review. The opportunity is a more programmable speech layer; the product is still a beta and should be treated that way.

Quick facts

Sources checked

Cohort context

Kalpa Labs is listed in Fall 2025. In our 2026-09-18 directory snapshot, 90 of 146 listed companies in that cohort have YC’s primary industry label B2B (61.6%). This is a current-directory comparison, not an original intake count or a performance ranking. Nine-cohort dataset.

Public website snapshot

Observed 2026-09-19 16:15 UTC in raw homepage HTML. This records visible metadata and advertised links, not agent execution or product quality.

Public observations · Collection method. Missing links here do not establish that a capability or file is absent elsewhere.