Problem

Voice AI labs need realistic, diverse conversational audio to train and evaluate models, but sourcing enough high-quality data across languages, accents, demographics, emotions, and real-world conditions is difficult. Because voice models depend heavily on training-data quality, the pain is operational as well as technical: teams must recruit contributors, collect recordings, transcribe and review them, and cover long-tail scenarios before they can reliably improve a model. The killer use case is supplying the data needed to train foundational voice models and fine-tune conversational AI agents.

Product / Service

Panels is a B2B audio-data platform that provides datasets from vetted contributors for speech and audio model development. Its offerings include multilingual speaker-separated audio, single-speaker scripted recordings across varied environments, and turn-taking evaluation data; it also designs bespoke datasets around a customer's requirements.

The delivery model combines research, in-house collection, transcription, quality assurance, and review. Panels first defines the target speakers, environments, use cases, edge cases, and acceptance criteria, then shares checkpoints and samples before scaling. It subsequently expands datasets based on model results and real-world feedback, helping customers improve coverage across accents, noise conditions, domains, and other failure modes without building the entire collection operation themselves.

Market

Panels competes in the B2B human-data and speech-data infrastructure market serving Voice AI and audio-model developers. Its adjacent competitive set includes broad AI-data providers such as Appen and Scale AI, as well as speech-focused providers such as Defined.ai. Panels appears more narrowly focused on high-quality conversational audio, vetted contributors, and iterative bespoke dataset design rather than offering a general-purpose annotation platform.

The company is an early-stage YC Summer 2025 startup listed as active, with Aaron Wenk as founder and a reported one-person team. Its launch materials claim a network of more than 10,000 vetted voice contributors spanning 20-plus languages and 100-plus countries, while its website says it works with frontier labs and early-stage startups. Public materials reviewed do not disclose revenue or named customers, so Panels' commercial scale and whether it is technically pre-revenue cannot be confirmed; the available evidence supports describing it as early-stage rather than as a company with demonstrated market-scale traction.