How a global automaker built automotive voice AI with a studio-quality in-car voice assistant dataset

Using highly controlled recording conditions, strict speaker validation rules and detailed demographic and acoustic metadata, we helped to build authentic in-cabin voice recognition.

TL;DR: automotive voice AI key takeways

• Delivered studio-quality voice data for a long-term automotive ASR program built around ultra-precise, clean in-cabin utterances.
• Collected 470 commands per speaker across 40 speakers per script, producing approximately 18.8K speech recordings per script plus calibration files, with 56.4K already delivered.
• Met a stringent 95% speaker validation threshold, with no more than 5 invalid files per 470-command speaker set accepted into delivery.
• Captured US English across 9 dialect groups, while maintaining diversity across age, gender and ethnicity requirements.
• Returned highly structured voice assistant training data with before-and-after calibration tones, speaker metadata and studio-grade audio specifications for vehicle testing.

Our customer: a multinational automaker manufacturer

Our customer is a leading global automaker developing a high-performance in-car voice assistant for vehicle environments where precision, consistency and broad speaker coverage matter. The goal was not simply to gather more audio, but to create highly controlled automotive voice AI data that could be used to test and improve how vehicle systems interpret spoken commands in realistic in-cabin scenarios.

Unlike general-purpose speech projects, this program required audio that could be replayed through a mannequin-based testing system inside a car. That meant every recording needed to be clean, repeatable and calibrated closely enough for the customer to simulate the original session conditions during testing.

Browse our speech and voice datasets

The challenge: why studio-quality voice data matters for an in-car voice assistant

Building a reliable in-car voice assistant requires more than generic speech samples. Automotive AI voice systems must be tuned against subtle differences in accent, speaker positioning, acoustic conditions like noisy environment speech and command delivery. This is especially true when those recordings are being replayed inside a vehicle test environment.

For this customer, the challenge was threefold:

• Quality over volume: each speaker recorded 470 commands, and the customer would only accept a speaker set if it met a 95% validation threshold with a maximum of 5 invalid files.
• Dialect coverage: each script had to include speakers representing 9 US dialect groups, with roughly 4–5 participants per dialect.
• Reproducible acoustics: every session required a calibration tone before and after recording so the customer could recreate the original audio conditions during internal testing.

That made this a high-risk, high-precision project. The customer was not looking for bulk audio collection. They needed domain-specific, studio-quality voice data that could stand up to strict validation and support downstream vehicle-system testing with minimal tolerance for error.

Explore our conversational AI solutions

Our capabilities: studio-grade speech data collection

Defined.ai was selected because the customer needed a partner that could manage specialized speech data collection across multiple US studios while maintaining consistent process control and transparent communication throughout the project.

Key capabilities that mattered most included:

• Multi-studio operational coverage: Defined.ai could recruit and record speakers across the US to meet precise dialect requirements.
• Studio-grade recording discipline: this project required controlled environments, calibrated equipment, repeatable microphone placement and strict monitoring of acoustic conditions for far-field speech.
• Detailed metadata capture: sessions included speaker demographics, dialect classification, seated mouth height, height and weight to support accurate in-vehicle mannequin testing.
• Transparent quality management: if a participant or file set did not meet requirements, Defined.ai worked with the customer to re-record, replace or correct submissions rather than letting issues accumulate downstream.

This combination of operational reach and quality control made Defined.ai well suited to collecting automotive speech recognition data where every file had to perform as intended in a testing environment, not just in a spreadsheet.

The solution: delivering high-fidelity automotive speech recognition data

Defined.ai delivered structured automotive speech recognition data tailored to the customer’s vehicle testing workflow.

1) Script-based speech collection at controlled scale

Each script contained hundreds of distinct automotive utterances, and every selected speaker recorded the full script. With dozens of unique speakers per script, that meant almost 20K speech recordings per script alone, excluding calibration files.

2) Studio-quality audio capture built for replay testing

Because the recordings would later be replayed through a mannequin inside a vehicle, the customer required clean, highly standardized audio. Delivered files followed studio-grade specifications, including:

• 48 kHz, 24-bit, single-channel WAV format
• Strict limits on leading and trailing silence
• Zero clipping tolerance
• Background noise levels capped at 50 dB(A) or below
• Before-and-after calibration tones recorded for each speaker session

This ensured the customer could reproduce the recording conditions more accurately during internal testing.

3) Diversity designed around dialect performance

For this project, variety was driven first by accent coverage. Each script had to represent a wide range of US dialect groups, including Western, Upper Midwestern, Midland, Mountain Southern, two Coastal Southern variants, Great Lakes, New York and New England. Within those constraints, Defined.ai also maximized diversity across ethnicity, age and gender.

4) Metadata and audio data annotation for genuine in-cabin simulation

Defined.ai collected more than audio. Each speaker session also returned audio data annotation that helped the customer reproduce in-vehicle testing conditions more precisely, including dialect, demographic information and seated mouth height. Even small differences in speaker positioning matter when audio is replayed through a mannequin into a vehicle cabin, so these details were part of the collection design from the start.

5) Long-term program delivery with quarter-by-quarter acceptance

This was not a one-off batch project: Defined.ai aligned to the customer’s specific cadence, delivering the level of quality needed for acceptance rather than optimizing only for collection speed.

Learn more about speech recognition datasets

The results: measurable progress in automotive voice AI data delivery

The customer received highly controlled automotive speech recognition data designed for vehicle-system testing and fine-tuning, not just broad model pretraining.

Measured and operational outcomes included:

• 56.4K studio-quality speech recordings delivered to date: 120 speakers total with 40 speakers per completed script plus approximately 240 calibration files captured in parallel.
• 9 US English dialect groups represented in every script to support more robust automotive voice AI tuning.
• 95% speaker validation threshold achieved for delivered speaker sets, with no more than 5 invalid recordings per 470-command submission submitted.
• 100% of completed scripts accepted.

Most importantly, the customer received voice assistant training data that could be replayed in its internal mannequin-vehicle testing setup with the consistency needed to evaluate and improve in-cabin understanding. For a use case where even small audio deviations can affect downstream results, that level of control was what mattered most.

What This Means for Automotive Voice AI

Defined.ai helped the customer build the acoustic and demographic foundation they needed through studio-quality voice data, without compromising on the precision needed for vehicle-grade validation. By combining controlled studio capture, detailed metadata and transparent, reliable delivery discipline, the project showed how the right data strategy can move automotive voice AI forward in a way that generic speech datasets cannot.

Looking to build or fine-tune an automotive AI project that needs accent coverage at scale, bespoke quality parameters or multilingual voice data? Talk to our automotive voice AI team