How CBB Labs enabled a dance studio to run 4× faster with AssemblyAI's Voice AI

Shivakami, director of Temple of Fine Arts in Texas, has been a classical dancer for over 30 years and has taught for more than a decade. Her studio serves close to 100 students across 13 courses, with students distributed globally.

All the operational work falls on her: class emails to every cohort, course setup, uploading practice videos individually, pausing and resuming student subscriptions, and managing monthly billing. Without dedicated administrative support, these tasks consume time she would otherwise spend with students.

DanceHall's dashboard provided visibility into student progress, attendance, and revenue through a modern SaaS interface. Text-based agentic actions added a layer of convenience, but interacting through chat still required thinking, typing, and waiting. In many cases, the time investment was comparable to clicking through the dashboard directly.

When voice agentic actions launched, the dynamic shifted. Instead of opening dashboards or composing prompts, Shivakami could ask questions and issue instructions verbally: "Who hasn't paid this month?" "Send reminders to the Tuesday class." "How many students attended last week?"

Bharatanatyam, one of the world's oldest continuously practiced classical dance traditions, originated in Tamil Nadu and draws from centuries of Sanskrit and South Indian musical traditions. Its technical vocabulary blends Tamil and Sanskrit terms. Teachers routinely use specialized words like thattadavu, nattadavu, tattimetti, alarippu, jatiswaram, and varnam in conversation. When speech-to-text mishandles these terms, the user must read the transcript, correct the error, and resume their flow. At that point, voice becomes a friction point rather than an efficiency gain.

CBB Labs evaluated several speech-to-text providers before selecting AssemblyAI for its ability to handle specialized terminology reliably. Shivakami can speak a full message in specialized vocabulary and receive an accurate transcription without stopping to re-correct. The accuracy continues improving as the system's knowledge base of her terminology grows. AssemblyAI handles the transcription layer while Sangs learns to interpret her world on top of it.

The result: she speaks, AssemblyAI transcribes accurately, and Sangs executes the intended action. Approximately 9 out of 10 voice requests now carry out the intended action correctly, making the system practical for daily use.

‍