Job Title: Audio Collection Specialists for Wake Word and Non-Wake Word Dataset Project

Project ID: SEP-0110(SH432)
Location: Remote
Contract Type: Freelance

About the Project

ServiceEdgePro is seeking experienced audio collection specialists to participate in our latest project focused on collecting wake words and non-wake words datasets across four languages: English (US), Spanish, Taiwanese Mandarin, and Japanese.

Project Scope

  • Collect wake words and utterances from a diverse group of speakers.
  • Each speaker will say the wake words 25 times to capture variations.
  • Participants will read 4 minutes of unique sentences for non-wake word recordings.

In Scope

  • Vendors with experience in audio collection.
  • Crowdsourcing teams.

Wake Words and Non-Wake Words Dataset Requirements

  • Wake Words: Each speaker must say the wake words 25 times. Variations include:

    • Hey Google
    • Hi Google
    • Hello Google
  • Speaker Demographics:

    • 500 speakers per language across various regions.
    • Gender distribution: 50% male, 50% female.
    • Age range: 20-60 years.
    • Males: 20-29 (20%), 30-39 (15%), 40-49 (10%), 50-59 (5%).
    • Females: 20-29 (20%), 30-39 (15%), 40-49 (10%), 50-59 (5%).
  • Languages:

    • 4 languages: English (US), Spanish, Taiwanese Mandarin, and Japanese.
    • Each language requires 12,500 utterances, totaling 50,000 utterances across all languages.

Audio Details

  • Format: WAV
  • Encoding: PCM, signed, 16-bit, little endian
  • Sampling Rate: 16 kHz
  • Bit Depth: 16 bits per sample
  • Channel: Mono (1 channel)

Recording Guidelines

  1. Peak amplitude should range between -21 dBFS and -6 dBFS.
  2. Record in a quiet environment with background noise <50 dBSPL. Ensure SNR >25 dB.
  3. No distortion, missing samples, or chopping.
  4. Normal speech speed, verified by subjective listening.
  5. Segment each wake word into separate .wav files, ensuring no part is cut off, with front/rear silence trimmed to <0.1 sec.

Non-Wake Word Recording

  • Each speaker must read 4 minutes of unique sentences, avoiding any selected wake words. Content should differ for each speaker and vary across the 4 languages.

Application Process

Interested candidates are invited to apply by filling out the Google Form below:

Apply Now