Hear from CIOs, CTOs, and other C-level and senior execs on data and AI strategies at the Future of Work Summit this January 12, 2022. Learn more

A preprint paper published by Fondazione Bruno Kessler (FBK) researchers proposes a benchmark — MuST-SHE — to evaluate whether speech translation systems fed textual data are constrained by the fact that sentences sometimes omit gender identity clues. The coauthors assert that these systems can and do exhibit gender bias, and that signals beyond text (like audio) provide contextual clues that might reduce this bias.

In machine translation, gender bias is at least partially attributable to the differences in how languages express female and male gender. Those with a grammatical system of gender, such as Romance languages, rely on a copious set of inflection and gender agreement devices that apply to individual parts of speech. That’s untrue of English, for instance, which is a “natural gender” language — it reflects distinction of sex only via pronouns, inherently gendered words (e.g., “boy,” “girl”), and marked nouns (“actor,” “actress”).

AI translation systems that fail to pick up on the nuances threaten to perpetuate under- or misrepresentation of demographic groups. Motivated by this, the researchers created MuST-SHE, a multilingual test set designed to uncover gender bias in speech translation.

MuST-SHE is a subset of TED talks comprising roughly 1,000 audio recordings, transcripts, and translations in English-French and English-Italian pairs from the open source MuST-C corpus, annotated with qualitatively differentiated and balanced gender-related phenomena. It’s subdivided into two categories:

  • Category 1: Samples where the necessary information to disambiguate gender can be recovered from the audio signal, when gender agreement depends only on the speaker’s gender.
  • Category 2: Samples where the disambiguating information can be recovered from the utterance content, where contextual hints such as gender-exclusive words (“mom”), pronouns (“she,” “his”), and proper nouns (“Paul”) inform about gender.

For each reference in the corpus, the researchers created a “wrong” one identical to the original except for the morphological signals conveying gender agreement. The result was a new set of references that are “wrong” compared with the correct ones in regard to the formal expression of gender, the idea being that the difference can be used to measure a speech recognition systems’ ability to handle gender phenomena.

In a series of experiments, the researchers created three speech recognition systems:

  • End2End, which was trained on the MuST-C and open source Librispeech data sets, augmented by automatically translating the original English transcripts into target languages.
  • Cascade, which shares the same core technology as End2End but which was trained on 70 million language pairs for English-Italian and 120 million for English-French from the OPUS repository, after which it was fine-tuned on MuST-C training data.
  • Cascade+Tag, a model identical to Cascade excepting tags added to the training data that indicate a speaker’s gender.

Interestingly, the researchers found that injecting gender information into Cascade didn’t have a measurable effect when evaluated on MuST-SHE. The difference values between the original and “wrong” references in the data set implied that all three systems were biased toward masculine forms.

When it came to the categories, Cascade performed the worst on Category 1 because it couldn’t access the speaker’s gender information it needed for a correct translation. End2End leveraged audio features to accurately translate gender, by contrast, but it exhibited the worst performance on Category 2 data — perhaps because it was trained on a fraction of the data used in Cascade and Cascade+Tag.

“If, like human beings, ‘machine learning is what it eats,’ the different ‘diet’ of machine translation and speech translation models can help them develop different skills,” wrote the researchers. “By ‘eating’ audio-text pairs, speech translation has a potential advantage: the possibility to infer speakers’ gender from input audio signals.”

The paper’s publication comes after Google introduced gender-specific translations in Google Translate chiefly to address gender bias. Scientists have proposed a range of approaches to mitigate and measure it, most recently with a leaderboard, challenge, and set of metrics dubbed StereoSet.


VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative technology and transact. Our site delivers essential information on data technologies and strategies to guide you as you lead your organizations. We invite you to become a member of our community, to access:
  • up-to-date information on the subjects of interest to you
  • our newsletters
  • gated thought-leader content and discounted access to our prized events, such as Transform 2021: Learn More
  • networking features, and more
Become a member