1{2 "document_id": "lb_qasper_07484380079dcec47eb573ab62eaff39a1a44b32006ec74e",3 "family": "public_qasper",4 "split": "test",5 "text": "Introduction\nIn recent years, gender has become a hot topic within the political, societal and research spheres. Numerous studies have been conducted in order to evaluate the presence of women in media, often revealing their under-representation, such as the Global Media Monitoring Project BIBREF0. In the French context, the CSA BIBREF1 produces a report on gender representation in media on a yearly basis. The 2017 report shows that women represent 40% of French media speakers, with a significant drop during high-audience hours (6:00-8:00pm) reaching a value of only 29%. Another large scale study confirmed this trend with an automatic analysis of gender in French audiovisuals streams, highlighting a huge variation across type of shows BIBREF2.\nBesides the social impact of gender representation, broadcast recordings are also a valuable source of data for the speech processing community. Indeed, automatic speech recognition (ASR) systems require large amount of annotated speech data to be efficiently trained, which leaves us facing the emerging concern about the fact that \"AI artifacts tend to reflect the goals, knowledge and experience of their creators\" BIBREF3. Since we know that women are under-represented in media and that the AI discipline has retained a male-oriented focus BIBREF4, we can legitimately wonder about the impact of using such data as a training set for ASR technologies. This concern is strengthened by the recent works uncovering gender bias in several natural language processing (NLP) tools such as BIBREF5, BIBREF6, BIBREF7, BIBREF8.\nIn this paper, we first highlight the importance of TV and radio broadcast as a source of data for ASR, and the potential impact it can have. We then perform a statistical analysis of gender representation in a data set composed of four state-of-the-art corpora of French broadcast, widely used within the speech community. Finally we question the impact of such a representation on the systems developed on this data, through the perspective of an ASR system.\nFrom gender representation in data to gender bias in AI ::: On the importance of data\nThe ever growing use of machine learning in science has been enabled by several progresses among which the exponential growth of data available. The quality of a system now depends mostly on the quality and quantity of the data it has been trained on. If it does not discard the importance of an appropriate architecture, it reaffirms the fact that rich and large corpora are a valuable resource. Corpora are research contributions which do not only allow to save and observe certain phenomena or validate a hypothesis or model, but are also a mandatory part of the technology development. This trend is notably observable within the NLP field, where industrial technologies, such as Apple, Amazon or Google vocal assistants now reach high performance level partly due to the amount of data possessed by these companies BIBREF9.\nSurprisingly, as data is said to be “the new oil\", few data sets are available for ASR systems. The best known are corpora like TIMIT BIBREF10, Switchboard BIBREF11 or Fisher BIBREF12 which date back to the early 1990s. The scarceness of available corpora is justified by the fact that gathering and annotating audio data is costly both in terms of money and time. Telephone conversations and broadcast recordings have been the primary source of spontaneous speech used. Out of all the 130 audio resources proposed by LDC to train automatic speech recognition systems in English, approximately 14% of them are based on broadcast news and conversation. For French speech technologies, four corpora containing radio and TV broadcast are the most widely used: ESTER1 BIBREF13, ESTER2 BIBREF14, ETAPE BIBREF15 and REPERE BIBREF16. These four corpora have been built alongside evaluation campaigns and are still, to our knowledge, the largest French ones of their type available to date.\nFrom gender representation in data to gender bias in AI ::: From data to bias\nThe gender issue has returned to the forefront of the media scene in recent years and with the emergence of AI technologies in our daily lives, gender bias has become a scientific topic that researchers are just beginning to address. Several studies revealed the existence of gender bias in AI technologies such as face recognition (GenderShades BIBREF17), NLP (word embeddings BIBREF5 and semantics BIBREF6) and machine translation (BIBREF18, BIBREF7). The impact of the training data used within these deep-learning algorithms is therefore questioned.\nBias can be found at different levels as pointed out by BIBREF19. BIBREF20 defines bias as a skew that produces a type of harm. She distinguishes two types of harms that are allocation harm and representation harm. The allocation harm occurs when a system is performing better or worse for a certain group while representational harm contributes to the perpetuation of stereotypes. Both types of harm are the results of bias in machine learning that often comes from the data systems are trained on. Disparities in representation in our social structures is captured and reflected by the training data, through statistical patterns. The GenderShades study is a striking example of what data disparity and lack of representation can produce: the authors tested several gender recognition modules used by facial recognition tools and found difference in error-rate as high as 34 percentage points between recognition of white male and black female faces. The scarce presence of women and colored people in training set resulted in bias in performance towards these two categories, with a strong intersectional bias. As written by BIBREF21 \"A data set may have many millions of pieces of data, but this does not mean it is random or representative. To make statistical claims about a data set, we need to know where data is coming from; it is similarly important to know and account for the weaknesses in that data.\" (p.668).\nRegarding ASR technology, little work has explored the presence of gender bias within the systems and no consensus has been reached. BIBREF22 found that speech recognizers perform better on female voice on a broadcast news and telephone corpus. They proposed several explanations to this observation, such as the larger presence of non-professional male speech in the broadcast data, implying a less prepared speech for these speakers or a more normative language and standard pronunciation for women linked to the traditional role of women in language acquisition and education. The same trend was observed by BIBREF23. More recently, BIBREF24 discovered a gender bias within YouTube's automatic captioning system but this bias was not observed in a second study evaluating Bing Speech system and YouTube Automatic Captions on a larger data set BIBREF8. However race and dialect bias were found. General American speakers and white speakers had the lowest error rate for both systems. If the better performance on General American speakers could be explained by the fact that they are all voice professionals, producing clear and articulated speech, but no explanation is provided for biases towards non-white speakers.\nGender bias in ASR technology is still an open research question as no clear answer has been reached so far. It seems that many parameters are to take into account to achieve a general agreement. As we established the importance of TV and radio broadcast as a source of data for ASR, and the potential impact it can have, the following content of this paper is structured as this: we first describe statistically the gender representation of a data set composed of four state-of-the-art corpora of French broadcast, widely used within the speech community, introducing the notion of speaker's role to refine our analysis in terms of voice professionalism. We then question the impact of such a representation on a ASR system trained on these data. BIBREF25\nMethodology\nThis section is organized as follows: we first present the data we are working on. In a second time we explain how we proceed to describe the gender representation in our corpus and introduce the notion of speaker's role. The third subsection introduces the ASR system and metrics used to evaluate gender bias in performance.\nMethodology ::: Data presentation\nOur data consists of two sets used to train and evaluate our automatic speech recognition system. Four major evaluation campaigns have enabled the creation of wide corpora of French broadcast speech: ESTER1 BIBREF13, ESTER2 BIBREF14, ETAPE BIBREF15 and REPERE BIBREF16. These four collections contain radio and/or TV broadcasts aired between 1998 and 2013 which are used by most academic researchers in ASR. Show duration varies between 10min and an hour. As years went by and speech processing research was progressing, the difficulty of the tasks augmented and the content of these evaluation corpora changed. ESTER1 and ESTER2 mainly contain prepared speech such as broadcast news, whereas ETAPE and REPERE consists also of debates and entertainment shows, spontaneous speech introducing more difficulty in its recognition.\nOur training set contains 27,085 speech utterances produced by 2,506 speakers, accounting for approximately 100 hours of speech. Our evaluation set contains 74,064 speech utterances produced by 1,268 speakers for a total of 70 hours of speech. Training data by show, medium and speech type is summarized in Table and evaluation data in Table . Evaluation data has a higher variety of shows with both prepared (P) and spontaneous (S) speech type (accented speech from African radio broadcast is also included in the evaluation set).\nMethodology ::: Methodology for descriptive analysis of gender representation in training data\nWe first describe the gender representation in training data. Gender representation is measured in terms of number of speakers, number of utterances (or speech turns), and turn lengths (descriptive statistics are given in Section SECREF16). Each speech turn was mapped to its speaker in order to associate it with a gender.\nAs pointed out by the CSA report BIBREF1, women presence tends to be marginal within the high-audience hours, showing that women are represented but less than men and within certain given conditions. It is clear that a small number of speakers is responsible for a large number of speech turns. Most of these speakers are journalists, politicians, presenters and such, who are representative of a show. Therefore, we introduce the notion of speaker's role to refine our exploration of gender disparity, following studies which quantified women's presence in terms of role. Within our work, we define the notion of speaker role by two criteria specifying the speaker's on-air presence, namely the number of speech turns and the cumulative duration of his or her speaking time in a show. Based on the available speech transcriptions and meta-data, we compute for each speaker the number of speech turns uttered as well as their total length. We then use the following criteria to define speaker's role: a speaker is considered as speaking often (respectively seldom) if he/she accumulates a total of turns higher (respectively lower) than 1% of the total number of speech turns in a given show. The same process is applied to identify speakers talking for a long period from those who do not. We end up with two salient roles called Anchors and Punctual speakers:\nthe Anchor speakers (A) are above the threshold of 1% for both criteria, meaning they are intervening often and for a long time thus holding an important place in interaction;\nthe Punctual speakers (PS) on the contrary are below the threshold of 1% for both the total number of turns and the total speech time.\nThese roles are defined at the show level. They could be roughly assimilated to the categorization “host/guest” in radio and TV shows. Anchors could be described as professional speakers, producing mostly prepared speech, whereas Punctual speakers are more likely to be “everyday people\". The concept of speaker's role makes sense at both sociological and technical levels. An Anchor speaker is more likely to be known from the audience (society), but he or she will also likely have a professional (clear) way of speaking (as mentioned by BIBREF22 and BIBREF8), as well as a high number of utterances, augmenting the amount of data available for a given gender category.\nMethodology ::: Gender bias evaluation procedure of an ASR system performance ::: ASR system\nThe ASR system used in this work is described in BIBREF25. It uses the KALDI toolkit BIBREF26, following a standard Kaldi recipe. The acoustic model is based on a hybrid HMM-DNN architecture and trained on the data summarized in Table . Acoustic training data correspond to 100h of non-spontaneous speech type (mostly broadcast news) coming from both radio and TV shows. A 5-gram language model is trained from several French corpora (3,323M words in total) using SRILM toolkit BIBREF27. The pronunciation model is developed using the lexical resource BDLEX BIBREF28 as well as automatic grapheme-to-phoneme (G2P) transcription to find pronunciation variants of our vocabulary (limited to 80K). It is important to re-specify here, for further analysis, that our Kaldi pipeline follows speaker adaptive training (SAT) where we train and decode using speaker adapted features (fMLLR-adapted features) in per-speaker mode. It is well known that speaker adaptation acts as an effective procedure to reduce mismatch between training and evaluation conditions BIBREF29, BIBREF26.\nMethodology ::: Gender bias evaluation procedure of an ASR system performance ::: Evaluation\nWord Error Rate (WER) is a common metric to evaluate ASR performance. It is measured as the sum of errors (insertions, deletions and substitutions) divided by the total number of words in the reference transcription. As we are investigating the impact on performance of speaker's gender and role, we computed the WER for each speaker at the episode (show occurrence) level. Analyzing at such granularity allows us to avoid large WER variation that could be observed at utterance level (especially for short speech turns) but also makes possible to get several WER values for a given speaker, one for each occurrence of a show in which he/she appears on. Speaker's gender was provided by the meta-data and role was obtained using the criteria from Section SECREF6 computed for each show. This enables us to analyze our results across gender and role categories which was done using Wilcoxon rank sum tests also called Mann-Whitney U test (with $\\alpha $= 0.001) BIBREF30. The choice of a Wilcoxon rank sum test and not the commonly used t-test is motivated by the non-normality of our data.\nResults ::: Descriptive analysis of gender representation in training data ::: Gender representation\nAs expected, we observe a disparity in terms of gender representation in our data (see Table ). Women represent 33.16% of the speakers, confirming the figures given by the GMMP report BIBREF0. However, it is worth noticing that women account for only 22.57% of the total speech time, which leads us to conclude that women also speak less than men.\nResults ::: Descriptive analysis of gender representation in training data ::: Speaker's role representation\nTable presents roles' representation in training data and shows that despite the small number of Anchor speakers in our data (3.79%), they nevertheless concentrate 35.71 % of the total speech time.\nResults ::: Descriptive analysis of gender representation in training data ::: Role and gender interaction\nWhen crossing both parameters, we can observe that the gender distribution is not constant throughout roles. Women represent 29.47% of the speakers within the Anchor category, even less than among the Punctual speakers. Their percentage of speech is also smaller. When calculating the average speech time uttered by a female Anchor, we obtain a value of 15.9 min against 25.2 min for a male Anchor, which suggests that even within the Anchor category men tend to speak more. This confirms the existence of gender disparities within French media. It corroborates with the analysis of the CSA BIBREF1, which shows that women were less present during high-audience hours. Our study shows that they are also less present in important roles. These results legitimate our initial questioning on the impact of gender balance on ASR performance trained on broadcast recordings.\nResults ::: Performance (WER) analysis on evaluation data ::: Impact of gender on WER\nAs explained in Section SECREF13, WER is the sum of errors divided by the number of words in the transcription reference. The higher the WER, the poorer the system performance. Our 70h evaluation data contains a large amount of spontaneous speech and is very challenging for the ASR system trained on prepared speech: we observe an overall average WER of 42.9% for women and 34.3% for men. This difference of WER between men and women is statistically significant (med(M) = 25%; med(F) = 29%; U = 709040; p-value < 0.001).\nHowever, when observing gender differences across shows, no clear trend can be identified, as shown in Figure FIGREF21. For shows like Africa1 Infos or La Place du Village, we find an average WER lower for women than for men, while the trend is reversed for shows such as Un Temps de Pauchon or Le Masque et la Plume. The disparity of the results depending on the show leads us to believe that other factors may be entangled within the observed phenomenon.\nResults ::: Performance (WER) analysis on evaluation data ::: Impact of role on WER\nSpeaker's role seems to have an impact on WER: we obtain an average WER of 30.8% for the Anchor speakers and 42.23% for the Punctual speakers. This difference is statistically significant with a p-value smaller than $10^{-14}$ (med(A) = 21%; med(P) = 31%; U = 540,430; p-value < 0.001) .\nResults ::: Performance (WER) analysis on evaluation data ::: Role and gender interaction\nFigure FIGREF25 presents the WER distribution (WER being obtained for each speaker in a show occurrence) according to the speaker's role and gender. It is worth noticing that the gender difference is only significant within the Punctual speakers group. The average WER is of 49.04% for the women and 38.56% for the men with a p-value smaller than $10^{-6}$ (med(F) = 39%; med(M) = 29%; U = 251,450; p-value < 0.001), whereas it is just a trend between male and female Anchors (med(F) = 21%; med(M) = 21%; U = 116,230; p-value = 0.173). This could be explained by the quantity of data available per speaker.\nResults ::: Performance (WER) analysis on evaluation data ::: Speech type as a third entangled factor?\nIn order to try to explain the observed variation in our results depending on shows and gender (Figure FIGREF21), we add the notion of speech type to shed some light on our results. BIBREF22 and BIBREF24 suggested that the speaker professionalism, associated with clear and hyper-articulated speech could be an explaining factor for better performance.\nBased on our categorization in prepared speech (mostly news reports) and spontaneous speech (mostly debates and entertainment shows), we cross this parameter in our performance analysis. As shown on Figure FIGREF26, these results confirm the inherent challenge of spontaneous speech compared to prepared speech. WER scores are similar between men and women when considering prepared speech (med(F) = 18%; med(M) = 21%; U = 217,160; p-value = 0.005) whereas they are worse for women (61.29%) than for men (46.51%) with p-value smaller than $10^{-14}$ for the spontaneous speech type (med(F) = 61%; med(M) = 37%; U = 153,580; p-value < 0.001).\nDiscussion\nWe find a clear disparity in terms of women presence and speech quantity in French media. Our data being recorded between 1998 and 2013, we can expect this disparity to be smaller on more recent broadcast recordings, especially since the French government displays efforts toward parity in media representation. One can also argue that even if our analysis was conducted on a large amount of data it does not reach the exhaustiveness of large-scale studies such as the one of BIBREF2. Nonetheless it does not affect the relevance of our findings, because if real-world gender representation might be more balanced today, these corpora are still used as training data for AI systems.\nThe performance difference across gender we observed corroborates (on a larger quantity and variety of language data produced by more than 2400 speakers) the results obtained by BIBREF24 on isolated words recognition. However the following study on read speech does not replicate these results. Yet a performance degradation is observed across dialect and race BIBREF8. BIBREF22 found lower WER for women than men on broadcast news and conversational telephone speech for both English and French. The authors suggest that gender stereotypes associated with women role in education and language acquisition induce a more normative elocution. We observed that the higher the degree of normativity of speech the smaller the gender difference. No significant gender bias is observed for prepared speech nor within the Anchor category. Even if we do not find similar results with lower WER for women than men, we obtained a median WER smaller for women on prepared speech and equal to the male median WER for the Anchor speakers.\nAnother explanation could be the use of adaptation within the pipeline. Most broadcast programs transcription systems have a speaker adaptation step within their decoding pipeline, which is the case for our system. An Anchor speaker intervening more often would have a larger quantity of data to realize such adaptation of the acoustic model. On the contrary, Punctual speakers who appear scarcely in the data are not provided with the same amount of adaptation data. Hence we can hypothesize that gender performance difference observed for Punctual speakers is due to the fact that female speech is further from the (initial non-adapted) acoustic model as it was trained on unbalanced data (as shown in Table ). Considering that Punctual speakers represent 92.78% of the speakers, this explains why gender difference is significant over our entire data set. A way to confirm our hypothesis would be to reproduce our analysis on WER values obtained without using speaker adapted features at the decoding step.\nWhen decoding prepared speech (hence similar to the training data), no significant difference is found in WER between men and women, revealing that the speaker adaptation step could be sufficient to reach same performance for both genders. But when decoding more spontaneous speech, there is a mismatch with the initial acoustic model (trained on prepared speech). Consequently, the speaker adaptation step might not be enough to recover good ASR performance, especially for women for whom less adaptation data is available (see Section 4.2.3).\nConclusion\nThis paper has investigated gender bias in ASR performance through the following research questions: i) what is the proportion of men and women in French radio and TV media data ? ii) what is the impact of the observed disparity on ASR performance ? iii) is this as simple as a problem of gender proportion in the training data or are other factors entangled ? Our contributions are the following:\nDescriptive analysis of the broadcast data used to train our ASR system confirms the already known disparity, where 65% of the speakers are men, speaking more than 75% of the time.\nWhen investigating WER scores according to gender, speaker's role and speech type, huge variations are observed. We conclude that gender is clearly a factor of variation in ASR performance, with a WER increase of 24% for women compared to men, exhibiting a clear gender bias.\nGender bias varies across speaker's role and speech spontaneity level. Performance for Punctual speakers respectively spontaneous speech seems to reinforce this gender bias with a WER increase of 27.2% respectively 31.8% between male and female speakers.\nWe found that an ASR system trained on unbalanced data regarding gender produces gender bias performance. Therefore, in order to create fair systems it is necessary to take into account the representation problems in society that are going to be encapsulated in the data. Understanding how women under-representation in broadcast data can lead to bias in ASR performances is the key to prevent re-implementing and reinforcing discrimination already existing in our societies. This is in line with the concept of “Fairness by Design\" proposed by BIBREF31.\nGender, race, religion, nationality are all characteristics that we deem unfair to classify on, and these ethical standpoints needs to be taken into account in systems' design. Characteristics that are not considered as relevant in a given task can be encapsulated in data nonetheless, and lead to bias performance. Being aware of the demographic skews our data set might contain is a first step to track the life cycle of a training data set and a necessary step to control the tools we develop.",6 "char_len": 25227,7 "document_hash": "17da848299284ab17e20a86e3c4e5e5933cdc0f11545427fdb2b156140b1bdb5",8 "questions": [9 {10 "question_id": "q00",11 "qtype": "qasper",12 "question": "How big is imbalance in analyzed corpora?"13 }14 ]15}