Human data for actions, outcomes, and what happens next.

Case study · Studio voice data

Expressive studio speech across US regional and South American accents

Off-the-shelf speech is read flat, cleaned up, and mostly one accent. The model needed real emotion, in the accents people actually have. Listen to clips from the studio sessions below.

01 Listen

Studio conversation clips

Six clips from three studio conversation sessions, cut at natural pauses. 48 kHz, 24-bit mono WAV, exactly as recorded: no compression, normalization, noise reduction, or editing.

  • 01

    17.5 s · −25.6 LUFS · −2.3 dBTP

  • 02

    15.6 s · −27.1 LUFS · −6.1 dBTP

  • 03

    12.5 s · −27.3 LUFS · −3.5 dBTP

  • 04

    21.2 s · −29.8 LUFS · −6.0 dBTP

  • 05

    20.2 s · −24.9 LUFS · −2.7 dBTP

  • 06

    16.1 s · −27.0 LUFS · −9.5 dBTP

02 Need

The model needed voices you can’t buy off the shelf

One line, many emotions, three deliveries:

“I told you not to open that door. Again: afraid, in a whisper.”

The line · the direction (underlined)

01

Gap

Existing speech is flat, processed, and mostly one accent.

02

Quality bar

Studio-clean, unprocessed, word-perfect.

03

Constraint

Speakers balanced by accent, age, and gender.

03 Solution

We moved quality control into the booth

  1. 01

    Record

    Studios record to the spec.

  2. 02

    Check

    Every take gets a verdict in seconds, so studios fix problems on the spot.

  3. 03

    Track

    Each speaker is followed to a complete set.

  4. →

    Delivered

    Accepted audio, metadata, and consent.

Fix it in the booth, not after the speaker goes home.

04 Acceptance

Every file is accented, expressive, and clean

RequirementStandard
US regionalNew York, East Coast, Southern, Midwestern, and more.
South AmericanEvery South American accent.
ExpressiveEmotions spoken, whispered, and shouted.
Raw audioNo noise reduction or processing.
AccurateScripts read exactly as written.
DocumentedMetadata and consent on every file.

What we deliver Accepted speech data with metadata, rights and provenance, and quality documentation.

Voice data, recorded to spec

What does your speech model need to hear?

Tell us the accents, emotions, speakers, and format you need. Contact Consequence Labs.

Scope a pilot