Ocular AI

Full-Duplex Conversational Datasets

Two-speaker conversations captured in full-duplex stereo across various languages and dialects. This feature preserves overlapping speech, backchannels, and natural disfluencies, providing a rich dataset for training conversational AI models. Users can explore and request samples from a diverse range of languages.

10hr Multi-Accent English ASR Dataset

An open-source dataset featuring multi-accent English speech recordings from 11 countries, balanced across genders and accents. With 7,377 recordings totaling 10.25 hours of audio, this dataset is designed for training Automatic Speech Recognition (ASR) models. Users can access the dataset to enhance their speech recognition capabilities.

Domain-Specific Speech Datasets

Task-anchored sessions across various scenarios such as medical intake, customer support, and emergency calls. These datasets are tagged by scenario, role, and intent, making them ideal for training vertical voice agents. Users can leverage these datasets to improve the performance of specialized AI applications.

Scripted Voice Datasets

Single-speaker performance reads from voice actors and trained narrators, featuring phonetically balanced content with controlled emotion ranges. This production-grade material is suitable for text-to-speech (TTS), voice cloning, and speech-to-speech applications. Users can utilize these datasets to create more natural-sounding AI voices.

Annotation & Evaluation Datasets

Comprehensive datasets that include word-level transcripts, diarization, prosodic markers, and continuous emotional tagging. These datasets provide the training signals necessary to turn raw audio into controllable and evaluable speech. Users can apply these datasets to enhance the evaluation and training of their AI models.

Data Foundry

A purpose-built engine that transforms raw human expertise into structured training data, alignment signals, and rigorous evaluations at scale. This feature allows users to capture the nuances of human expertise across various domains, ensuring that AI models are trained with real-world knowledge. Users can interact with the Data Foundry to create tailored datasets for their specific needs.

Elite Expert Network

A network of domain experts, linguists, and researchers who collaborate to capture and encode human expertise into training data. This feature ensures that the datasets reflect the complexity of human reasoning and interaction, providing a rich foundation for AI training. Users can leverage this network to access high-quality, expert-validated datasets.

Full-Duplex Captures

Captures two-speaker conversations at 48 kHz with isolated channels, preserving overlap, backchannels, and barge-in interactions verbatim. This feature provides the training audio necessary for developing real-time conversational voice agents. Users can utilize these captures to enhance the naturalness and responsiveness of their AI systems.

Emotional Tagging

A feature that adds emotional context to audio datasets, allowing for the training of AI models that can recognize and respond to human emotions. This capability enhances the emotional intelligence of voice interactions, making AI systems more relatable and effective in communication. Users can apply emotional tagging to improve user experience in conversational AI.

Prosodic Markers

Incorporates prosodic features into audio datasets, capturing the rhythm, stress, and intonation of speech. This feature is essential for training AI models that need to understand and generate natural-sounding speech. Users can leverage prosodic markers to enhance the quality of voice synthesis and recognition.