Improve Your Speech Recognition Models

Build or train your speech models in specific languages and domains or expand the scope with our high quality AI training datasets. Our expertise includes virtual or voice assistants, ASR, STT or TTS engines, call center IVR systems or vehicle infotainment assistants.

Browse Catalog

Monologue Speech Collection

Collect single speaker scripted, guided or spontaneous speech datasets, in broadband or narrowband.

Dialogue Speech Collection

Collect Agent and Caller or Caller and Bot interactions in guided or spontaneous speech datasets.

Speech-to-Text Transcription

Our transcription workflows provide data collection, correction and validations to improve your STT system.

Speech Validation

Speech data is validated with our certified crowd incorporating inter-annotator agreements and gold sets.

Speech Quality Guarantee

Speech recognition systems require the highest quality AI training data to perform properly, otherwise, it  will frustrate rather than delight.   Our speech collection, transcription and validation workflows utilize a variety of ML algorithms and crowd quality checks that allow us to guarantee our quality.

Some of our quality metrics include:

Word Error Rate

Speech dataset guarantee <5% for single speaker and <10% for multiple speakers.

Signal-to-Noise Ratio

Controls dataset variation in background noise, ambient sounds, and other audio.

Nativeness

Ensures the datasets use native speakers for each language.

Text-Audio Match

Human in the loop transcription validations check for exact matches.

Success Stories

Mastercard’s R&D Labs needed unique, multi-lingual text data that covered 20 designated payment scenarios in English and Spanish, and they needed it fast.

Keeping a nation’s lights on means constantly inspecting electricity poles for damage. EDP partnered with DefinedCrowd to improve Asset Performance Management processes.

With the rise of voice technology, this leading global provider of audio equipment wanted to develop an automatic speech recognition (ASR) model.

A global electronics maker came to DefinedCrowd with the goal of building more inclusive facial recognition models, requiring accurately annotated images with highly specific criteria.

Smart companies see the pile of unstructured text floating through the digital realm as a strategic goldmine of consumer insights.

A Fortune 500 Tech company needed comprehensive speech training data in French that accounted for a wide range of dialects, requiring diverse data in terms of age, gender and regional dialects.

A visionary Fortune 500 Tech company leveraged sentiment analysis models to dig beyond surface-level understandings to extract granular-level insights.