Ocular AI is an applied research lab dedicated to building data infrastructure and a network of human experts to encode real-world knowledge into frontier AI models, spanning multiple modalities, languages, and domains. At its core is Data Foundry—a purpose-built engine that transforms raw professional expertise into structured training data, alignment signals, and rigorous evaluation at scale. The team works with elite professionals to capture how they actually work, including the reasoning behind diagnoses, the intuition in negotiations, the rhythm of native speakers, the engineering insight in design decisions, and the micro-expressions machines have never been taught to recognize. From PhD mathematicians and voice actors to constitutional lawyers and linguists, every discipline, accent, and dialect has its own dedicated data pipeline.
Key capabilities:
- Converse benchmarks: including Converse-STT (conversational speech-to-text benchmark) and Converse-Diarization (conversational speaker diarization benchmark), using Ocular AI's open-source testing framework to evaluate frontier AI models on real-world two-person American English conversations. The STT benchmark is measured by word error rate (WER) and character error rate (CER), and the Diarization benchmark is measured by diarization error rate (DER).
- Benchmark results: Converse-STT v2 covers 19 models, with Scribe v2 (ElevenLabs) ranking first at 1.22% WER, Grok Voice Transcribe 2.0 (xAI) at 2.94%, Gemini 3.5 Transcribe (Google) at 3.13%, and so on. Converse-Diarization covers 11 models, with Pyannote Community-1 ranking first at 27.98% DER, DiariZen WavLM Large s80 MD v2 (BUT) at 28.16%, Nemotron 3 Diarization (NVIDIA) at 28.21%, and so on.
- Conversational AI datasets: provides full-duplex conversation datasets (48 kHz two-channel recordings, preserving overlap, backchannel, and interruptions), domain-specific speech datasets (scenarios such as healthcare, customer service, technical interviews, and emergency calls), scripted speech datasets (voice actors and trained narrators, phonetically balanced, controlled emotional range, multi-version recordings), and annotation and evaluation datasets (word-level transcription, speaker diarization, prosodic markers, scene and role labels, continuous emotion annotation, human preference ratings), supporting more than 40 languages, including American English.
How to use: Ocular AI primarily serves frontier AI labs and Fortune 100 enterprises, providing data infrastructure and an expert network. Specific usage details are not yet provided in detailed public documentation or a self-service registration process.
Applicable scenarios: AI speech model development requiring high-quality, real-world conversational training data, including real-time conversational voice agents, vertical-domain voice agents, TTS, voice cloning, speech-to-speech, and more. It is also suitable for researchers conducting benchmark evaluations of speech AI models.
Limitations: Official materials do not clearly disclose information such as pricing, specific usage steps, or platform support details, and none is currently provided.Continuous emotion annotation, human preference ratings), supporting more than 40 languages, including American English.
How to use: Ocular AI primarily serves frontier AI labs and Fortune 100 enterprises, providing data infrastructure and an expert network. Specific usage details are not yet provided in detailed public documentation or a self-service registration process.
Applicable scenarios: AI speech model development requiring high-quality, real-world conversational training data, including real-time conversational voice agents, vertical-domain voice agents, TTS, voice cloning, speech-to-speech, and more. It is also suitable for researchers conducting benchmark evaluations of speech AI models.
Limitations: Official materials do not clearly disclose information such as pricing, specific usage steps, or platform support details, and none is currently provided.
