Digital assistant voice

Command translation

Challenge

Improve recognition of non-English voice commands by improving the naturality of translations used in bot training.

Our approach

Removed wake words from client-supplied source text prior to translation. 

Sourced high-quality native translators with bidirectional expertise in US English and native language modern vernacular, as well as high scores from previous human translation assignments. 

Used proprietary MT Detection tool to measure the distance between translator-supplied strings and known machine-translated strings, automatically flagging suspicious strings for human review and retranslation as appropriate. 

Validated that strings bore no trace of machine translation or post-editing, and performed review, proofreading, and auto QA checks for punctuation and grammar.

Results

Extremely high-naturality translation datasets that are objectively validated to comprise human translation in their entirety.

Project details

Customer profile

Global retail and cloud computing giant

Data Type

Speech utterances Translation of user voice commands in text format

Audio Transcription & Annotation

Audio Categorization

Scope of Work

5-10 translators and 2-3 reviewers

35,000 words per language pair Local language to English

Geographical Coverage

English to 15 languages

Collection Methodology

Remote in particular country

Duration of Engagement

Ongoing work with batch turnaround time (TAT) within 2 weeks