CaptureLabz Recording Protocol v1.0

CaptureLabz Research Program

CaptureLabz is being developed as an applied research program within the CaptureLabz voice dataset initiative. The program investigates how structured speech datasets affect the reliability, robustness, and governance of modern speech recognition and language technology systems. Its work focuses on dataset methodology, acoustic-condition variation, and the development of speech resources for languages that remain underrepresented in current AI training pipelines.

• CaptureLabz Overview
• Recording Protocol
• Speech Robustness Benchmark
• Nahuatl & Mixtec Pilot Study

Primary Research Objectives

• Evaluate how structured multi-condition speech datasets influence ASR robustness
• Develop reproducible dataset capture methods for low-resource languages
• Document governance practices for ethically sourced voice data
• Create dataset structures suitable for AI training, benchmarking, and language preservation

Current Research Components

• Recording Methodology — controlled multi-condition speech capture
• Robustness Benchmark — experiments evaluating model performance under acoustic variation
• Language Dataset Pilots — initial structured datasets for Nahuatl and Mixtec

These components are documented through the CaptureLabz research documentation within XPGuess Learn and are intended to support both academic research and applied speech technology development.

What CaptureLabz Is

CaptureLabz is not just a recording interface. It is a standardized speech dataset protocol and benchmark framework intended to help communities, researchers, and technical teams create more useful voice data for AI systems.

Most low-resource language efforts focus on preservation alone or collect audio in one narrow recording condition. CaptureLabz introduces a repeatable structure that combines controlled acoustic variation, documented metadata, and governance-aware collection logic.

Why This Protocol Exists

Modern speech recognition systems are often trained on audio collected in quiet environments with limited acoustic diversity. That creates a reliability problem. Models may perform well in clean settings, then degrade when speech is recorded from farther away, in a different room, or under more natural listening conditions.

CaptureLabz exists to reduce that gap. It introduces a structured protocol that allows speech datasets to be created with known acoustic variation rather than accidental noise alone.

For low-resource languages, this matters even more. Many languages do not have enough recorded material to support multiple rounds of trial-and-error data collection. The protocol is intended to make each session more valuable from the beginning.

Core Recording Conditions

The first version of the CaptureLabz protocol uses four primary recording conditions. These are designed to create a small but meaningful matrix of acoustic diversity that can be reproduced across languages and communities.

Additional conditions can be layered later, including device variation, outdoor recording, speaker movement, or bilingual crossover patterns. Version 1.0 begins with a compact set that is simple enough to reproduce and strong enough to support benchmark testing.

Dataset Structure

CaptureLabz datasets are designed as structured releases rather than loose folders of audio. A complete release should include audio files, transcripts, metadata, and condition labels that allow researchers to reproduce training and evaluation experiments.

Core fields

speaker_id
language
dialect
recording_condition
environment
device_type
transcript
ipa
audio_file
session_id
consent_status
qa_status

This structure allows researchers to separate by condition, compare training regimes, and evaluate whether performance changes are tied to recording distance, pacing, or dialect tagging rather than to undocumented variation.

Example release logic

capturelabz/
├── audio/
│   ├── nahuatl/
│   └── mixtec/
├── metadata/
│   ├── speakers.json
│   ├── sessions.json
│   └── dialect_tags.json
├── transcripts/
│   ├── nahuatl_orthography.csv
│   ├── nahuatl_ipa.csv
│   ├── mixtec_orthography.csv
│   └── mixtec_ipa.csv
└── tools/
    ├── loader.py
    ├── train_test_split.json
    └── benchmark_notes.md

Speech Robustness Benchmark

CaptureLabz is intended to function as more than a dataset release process. It is also a speech robustness benchmark for testing how models perform under acoustic distribution shift.

A simple benchmark question looks like this:

Example benchmark setup

• Train baseline model on Close / Normal recordings only
• Train intervention model on all four CaptureLabz conditions
• Evaluate both on unseen distance and natural-environment recordings
• Measure differences in error rate and degradation under shift

Typical outputs may include Word Error Rate (WER), Character Error Rate (CER), performance decay by condition, and condition-aware model comparisons.

Pilot Languages and Initial Use

The early CaptureLabz pilot is being shaped around Nahuatl and Mixtec, including speaker access through existing trusted community relationships in Mexico. These languages are important not only because they are historically and culturally significant, but because they are also underrepresented in modern AI speech systems.

The pilot is intended to validate:

• that the multi-condition recording protocol is practical in real community settings,
• that the metadata structure is usable for research packaging, and
• that acoustic diversity improves robustness compared with narrow collection methods.

Future pilots may expand into additional languages and regions once the first release and benchmark logic are stabilized.

Why Researchers Care

Researchers do not only need more data. They need better-controlled data, clearer provenance, and datasets that expose where models fail. CaptureLabz is designed to support that by producing speech data with explicit condition labels and a documented protocol.

Terminology, Frameworks, and Foundational Work

XPGuess — Extended Performance Guessing — is an educational decision-learning construct used to explore how development paths and outcomes unfold over time.

Natural Technical Governance (NTG) documents training and participation using first principles rather than subjective opinion.

The conceptual foundations derive from earlier technical work by Michael A. Piña, including biomechanical and developmental research.

Reference: “Beginning and Staying with the Basics: Building from the Ground Up”

Additional work: Coach Teaches Animals: Gymnastics Stretching