Humyn Labs

Humyn Labs is an artificial-intelligence company founded in 2026 by co-founders Manish Agarwal and Ishank Gupta. Headquartered in San Francisco, it develops training data and data infrastructure for robotics, physical AI, and voice AI, capturing human activity in real-world settings as synchronized, provenance-tracked sensory recordings.^[1]​^[3]​^[2]​

Overview

Humyn Labs describes its platform as a way to address the shortage of structured real-world recordings for physical AI. Its examples include recordings of household cleaning and manual vehicle assembly, activities it says are not represented in the same way as the internet-derived material commonly used to train other AI systems.^[2]

The company says its recording system synchronizes four sensor streams during capture: camera, inertial measurement unit (IMU), audio, and kinematics. It identifies contributor ID, environment, timestamp, and hardware serial as provenance fields, and separately says that contributor verification occurs on-chain. Humyn Labs describes the accompanying workflow as covering synchronization, validation, processing, quality control, and benchmark evaluation, with deliverables in formats including MCAP, RLDS, and LeRobot v3.^[2]

The company is developing tools for sourcing, validating, processing, and labeling episodic data. It characterizes its data collection as “egocentric” and “source-first”: recordings capture first-person human activity and how people see, navigate, and interact with their surroundings.^[2][3]

Humyn Labs says its expert network extends across more than 20 countries, including Vietnam, Nigeria, Egypt, South Korea, India, and Brazil. It presents this geographic spread as a way to provide models with exposure to different people, environments, and practices, particularly in regions it says are underrepresented in Western-trained models.^[2][4]

Data Collection and Quality Control

Humyn Labs describes its general data workflow as a sequence of requirements definition, workforce sourcing and screening, collection, quality control, technical processing, and dataset delivery. Requirements can include data specifications, expertise, volume, timelines, and quality thresholds. The company says its workflows track provenance from the initial request through the delivered dataset.^[8]

The company distinguishes its verified contributors and domain experts from anonymous crowdsourced work. It says contributors are selected for demonstrated skills and domain experience, and that layered quality checks are intended to reduce data-quality risks and provide records that clients can use to substantiate datasets to stakeholders.^[8]

Humyn Labs lists examples of completed or ongoing work that include a dataset of 120,000 faces across five regions for facial recognition, more than 1,000 hours of production-scale text-to-speech data for India’s banking, financial services, and insurance sector, and 50,000 hours of speech data for 10 languages.^[8]

The company’s physical-AI materials report 10,000 hours of visual data from more than 20 countries. The sight offering includes six-degree-of-freedom pose, hand-skeletal data, IMU and simultaneous localization and mapping (SLAM) data, and episode-level annotation; Humyn Labs says it discards less than 15 percent of this data. Its mobility offering includes seven-degree-of-freedom arm kinematics, whole-body pose, joint torque, and spatial mapping, captured through the same visual-data pipeline. Touch data is an emerging area involving force, resistance, and compliance measurements from sensor-equipped teleoperated grippers during manipulation-data collection.^[2][1]

History

Humyn Labs says it was founded in response to AI teams treating training data as a commodity, which it associates with anonymous crowds, inconsistent standards, and limited traceability. Its stated mission is to bridge human and artificial intelligence through verified human expertise, fair compensation, and ethically sourced data.^[4]​

Co-founders Manish Agarwal and Ishank Gupta lead the company. Humyn Labs’ background materials describe one founder as having led and scaled Nazara Technologies as chief executive to what the company characterizes as India’s first gaming initial public offering, at a peak valuation exceeding $1 billion. The same materials say the other scaled AB InBev Central and South Asia to more than $150 million and built consumer businesses across India, China, the United Kingdom, and South Africa.^[5]​^[3]​^[4]​

On April 13, 2026, Humyn Labs announced a $20 million commitment to expand data-collection operations and voice capabilities, with planned activity in India, Southeast Asia, Latin America, and West Asia. The company said the investment would support data for training robots and voice models.^[3]​^[1]​

The company also announced plans to establish Robotics Labs to build high-fidelity simulation environments and world models incorporating real-world data into training frameworks. Humyn Labs is a member of NVIDIA’s Inception program.^[3]​^[1]​

Products and Services

Physical AI and Robotics Data

Humyn Labs offers collection and annotation of real-world sensor data for robotics and other physical AI systems. The company argues that simulation can miss rare situations, edge cases, and variation in real-world physics, while field collection of synchronized sensor data can be costly and logistically demanding.^[6]​

Its field teams collect data in real environments using sensor configurations that can include multi-camera arrays, radar, IMUs and other motion sensors, GPS or global navigation satellite system (GNSS) equipment, and ultrasonic sensors. Supported data includes multi-camera video, motion and odometry records, and robotic manipulation such as pick-and-place, grasping, and assembly.^[6]​

Annotation services include 3D bounding boxes, point-cloud semantic segmentation, object tracking across frames, scene classification and tagging, and sensor-calibration validation. Humyn Labs says its annotators have experience in autonomy, robotics, or spatial computing; it also describes them as domain experts with on-chain reputations.^[6]​

The company’s described workflow begins with a client specifying sensor configurations, environments, edge cases, and annotation schemas. Field teams then collect synchronized recordings, which domain experts annotate; datasets go through peer review and centralized quality review. Humyn Labs says safety-critical work, such as autonomous-driving data, receives an additional domain-expert review, and that deliverables include annotation provenance and inter-annotator agreement metrics.^[6]​

Datasets can be delivered in formats such as KITTI, nuScenes, Waymo Open Dataset, COCO 3D, or a custom schema, with calibration files and timestamp-synchronization metadata. The company identifies robotics, industrial automation, autonomous vehicles, and drones as applications, including factory-floor inspection, pick-and-place systems, and autonomous forklifts.^[6]​

Voice Data

Humyn Labs collects and annotates voice data for automatic speech recognition (ASR), text-to-speech (TTS), voice cloning, and speech synthesis. Its service covers more than 50 languages, including major global and low-resource languages, and it lists Indic-language coverage such as Hindi, Tamil, Telugu, Kannada, Malayalam, Bengali, and Marathi. Recordings can be configured for particular accents, dialects, demographic groups, and multilingual code-switching.^[7]​

Recording types include scripted passages and phonetically balanced sentences, spontaneous speech and multi-turn conversations, emotional speech, wake-word and command recordings, multi-speaker dialogue, speech in noisy settings, and varied speaking styles such as whispers and shouts. The company offers audio in configurable sample rates of 16–48 kHz and 16

• or 24-bit depth, with WAV, FLAC, or MP3 delivery.^[7]​

The voice-data workflow starts with a client specification covering language, accent, demographics, audio format, utterance types, and volume. Humyn Labs says it verifies speaker identity and demographic information before recording, then checks audio quality and transcript alignment through automated and human review. Its automated checks include noise floor, clipping, silence ratio, and sample rate; human review covers transcript accuracy and speaker identity. The company says recordings that fail its requirements are rejected before delivery.^[7]​

Humyn Labs says speakers provide verified informed consent and that data handling follows the General Data Protection Regulation (GDPR) and regional privacy requirements. It also offers annotation of existing audio, including transcription, speaker labels, timestamps, and emotion tags, alongside new recording collection.^[7]​

The company reports 50,000 hours of sound data across 33 languages and says this collection is revenue-generating and has been evaluated using its BRIDGE benchmark. Its voice-data service page describes configurable collections at different scales, with pilot projects of 50–100 hours typically taking two to four weeks and projects exceeding 1,000 hours across multiple languages taking two to six months.^[2]​^[7]​

BRIDGE Speech Benchmarks

On May 13, 2026, Humyn Labs published BRIDGE, short for Benchmark of Regional & International Data for Global Evaluation. It evaluates commercial speech-recognition systems using real conversational audio in non-English languages across the Global South. Co-founders Manish Agarwal and Ishank Gupta were identified with the project.^[5]​

The benchmark evaluated 15 commercial models across 22 languages, including Latin American Spanish dialects, Brazilian Portuguese, and Vietnamese. Humyn Labs said the languages covered are spoken by more than 5.5 billion people. Its tests used real two-person conversations collected across multiple regions rather than scripted or internet-scraped audio, and the dataset was made available on Hugging Face.^[5]​

BRIDGE used seven evaluation metrics, including Semantic Similarity, Code-Switch F1, Loan Word Word Error Rate (WER), Phoneme-Informed Error Rate, and Word Information Lost. The report argued that rankings based on a single aggregate score can obscure differences in regional performance and types of recognition errors.^[5]​

The report ranked ElevenLabs Scribe v2 first overall, with a 10.6% WER, but third in Vietnamese; AssemblyAI Universal performed best in Vietnamese despite ranking 12th overall. It also reported substantial performance gaps for Venezuelan Spanish, Brazilian Portuguese, and cross-regional speaker pairs, while finding that models with high Vietnamese WER nevertheless preserved more than 93% of meaning accuracy.^[5]​

In September 2026, Humyn Labs announced BRIDGE ASR 2.0, which it described as an independent global benchmark evaluating 23 models across six metrics. Its materials emphasized that a single WER score can conceal different model failure modes and variation related to language and pause length. BITS Pilani researcher Divyajot Singh analyzed raw BRIDGE ASR 2.0 data and reported findings beyond the leaderboard scores.^[1]​

Agarwal said that English-first, internet-trained benchmarks did not provide independent validation representative of how Global South users speak, despite their use in commercial deployment decisions. Gupta argued that evaluation systems based on English phonology were not sufficient for comparing non-English speech, and that a single aggregate score could not support decisions across regions.^[5]​

Organization and Ecosystem

Humyn Labs is headquartered in San Francisco and has a workforce distributed across India, Brazil, and the United States.^[1] The company identifies founding member Anoushka Ananth, co-founder Ishank Gupta, and technical program manager Itisha Dubey among its personnel; it says Dubey built a naturalness-detection layer to screen collected audio clips for suitability as training data.^[1]