Automotive Wake Word Data Collection Across Four Languages and Accents

LXT ran a multilingual wake word data collection program across four locales, delivering accent-authentic, gender-balanced speech data for an automotive voice assistant.

The Challenge

A leading speech recognition technology vendor needed multilingual wake word data to expand training coverage for an automotive voice assistant deployed across multiple vehicle brands. The wake word detection system had to respond reliably to five branded phrases spoken by drivers across distinct regions, each with their own phonetic patterns, accent variation, and linguistic norms.

Specifically, the program had to:

• Source speakers across four target locales: Mexican Spanish, Japanese, Canadian French, and Gulf Arabic.
• Capture region-authentic accent variation for the English wake phrases, which had to be pronounced with the accent of the respective target region.
• Handle mixed-language utterances, where the wake phrase was spoken in English and the remainder in the target language.
• Recruit gender-balanced contributor pools of 120 to 200 speakers per locale within defined delivery windows.
• Execute recording collection entirely on the customer’s proprietary app, with assets transferred via links provided by email.

Off-the-shelf speech datasets did not cover the required locale and accent combinations for automotive-grade wake word detection. In-house audio data collection would have required the vendor to build regional recruitment infrastructure across four markets simultaneously, which was neither operationally feasible nor cost-efficient within the project timeline.

The Solution

LXT designed an automotive wake word data collection program that used its proprietary contributor platform to source and recruit speakers across all four locales. Once enrolled, contributors completed their recording sessions on the customer’s own web platform, keeping the data pipeline under the vendor’s control from the outset.

Accent-Authentic Wake Phrase Capture
Contributors were required to pronounce five branded automotive wake phrases in English with an accent representative of their target region. Audio reference examples were provided to each participant so they could calibrate their pronunciation before recording. This approach captured the natural acoustic variation a deployed system would encounter from real drivers.

Mixed-Language Utterance Design
Each recording session combined an English wake phrase with a continuation spoken in the target language. This matched the real-world interaction pattern of drivers who switch between languages mid-sentence, providing training data that reflected genuine in-vehicle usage rather than laboratory conditions.

Demographic Balance, by Design
Contributor pools were structured for gender balance across all four locales. This was built into the recruitment brief from the outset, ensuring the resulting dataset reflected the diversity of the actual driver population rather than relying on post-hoc rebalancing.

In-Session Quality Gate
The customer’s platform included an automatic loudness check on every take. Contributors received immediate feedback and could re-record any item that did not pass before advancing, keeping session quality consistent without a separate manual review pass. An identical three-take audio quality check ran during enrollment so that speakers with unsuitable recording conditions were identified before entering the main session.

Contributor Metadata and Consent
Each contributor completed a structured enrollment flow: EULA acceptance, demographic self-declaration (gender, age, native and second language, childhood region, current region), and an audio quality gate. This produced the per-speaker metadata needed to verify demographic balance and regional authenticity across the dataset, and ensured all contributors provided informed consent prior to recording.

Platform-Integrated Delivery
All recordings were performed directly on the customer’s web platform, eliminating format conversion overhead and keeping the data pipeline under the vendor’s control. Asset links were distributed by email, and LXT managed contributor coordination and throughput to meet per-locale delivery schedules.

Project Specifications

The exact numbers and rules behind the approach above, scannable at a glance.

• Wake Phrases: 5 branded automotive wake phrases (confidential)
• Wake Phrase Language: English, pronounced with target-region accent
• Continuation Language: Target locale language
• Locales: Mexican Spanish; Japanese, Canadian French, Gulf Arabic
• Speakers — Mexican Spanish: 150 speakers · 57 recordings/speaker
• Speakers — Japanese: 120 speakers · 153 recordings/speaker
• Speakers — Canadian French: 120 speakers · 121 recordings/speaker
• Speakers — Gulf Arabic: 200 speakers · 153 recordings/speaker
• Total Speakers: 590 (gender-balanced across all locales)
• Estimated Total Recordings: ~72,000 (sum of per-locale speaker × recording targets)
• In-Session Loudness Check: Automatic per-recording; contributors re-recorded items that did not pass before advancing
• Enrollment Audio Quality Check: 3-take loudness check before main session access
• Demographic Metadata Collected: Gender, age, native language, second language, childhood, region, current region
• Quality Assurance: Automated loudness gate on customer’s platform
• Reference Material: Audio reference examples for all five wake phrases embedded directly in the recording interface
• Delivery Platform: Customer’s proprietary web app; asset links distributed by email

Contributor Task Flow

1. Receive enrollment link via email and open the customer’s web-based recording platform.
2. Accept the end-user license agreement, confirming consent for the recordings to be used for speech recognition research and model training.
3. Self-declare demographic information: gender, age, native language, second language, childhood region (ages 2–15), and current region.
4. Complete an audio quality check: record a test phrase three times in a quiet environment; the platform automatically evaluated each take for loudness before granting access to the main session.
5. Listen to the embedded reference audio for each of the five wake phrases to calibrate pronunciation and accent before starting.
6. Record each prompted phrase one at a time; the platform evaluated each recording for loudness automatically and prompted a re-record if the take did not pass.
7. Complete the full quota of recordings for the session (up to 153 recordings per speaker depending on locale).
8. Submit the completed session to receive a speaker ID confirmation.

Outcomes

• Delivered accent-authentic, gender-balanced wake-phrase recordings across four linguistically and phonetically distinct locales.
• Provided training data that reflected real-world mixed-language driver behaviour, combining English wake phrases with target-language continuations.
• Met per-locale delivery schedules across a multi-contract engagement spanning two separate purchase orders.
• Executed the full collection pipeline on the customer’s existing app infrastructure.
• Executed the program across four linguistically distinct locales, meeting per-locale delivery schedules throughout the engagement.

Why LXT

• Regional contributor networks for multilingual wake word data collection across all four target locales (Mexican Spanish, Japanese, Canadian French, Gulf Arabic), with the capacity to recruit to gender-balance requirements within project timelines.
• Mixed-language expertise: experience managing audio data collection tasks where phonetic accuracy in the wake word is as important as fluency in the continuation language.
• Flexible platform integration: LXT coordinated contributor workflows on the customer’s own app, adapting to the vendor’s existing data pipeline rather than requiring a proprietary tool.
• Proven scalability: demonstrated ability to expand automotive wake word data collection into adjacent locales at any time if needed.