アフリカ系アメリカ人風のTTSの作成：ガイドライン、技術的課題、驚くべき評価

Core Concepts

アフリカ系アメリカ人女性の専門的な音声を再現するために開発されたTTSシステムにおける人種表現の課題と結果を探求する。

Abstract

この論文は、AIエージェントやロボットの表現が主に白人であることに焦点を当て、アフリカ系アメリカ人女性の専門的な音声を再現するTTSシステムの開発プロセスと技術的課題について探究しています。以下は論文内容の概要です。
1. 導入

現在使用されている米国英語TTSシステムは主に白人として認識されます。
プロジェクトの最終目標は、デザイナーや企業が明確にアフリカ系アメリカ人として認識できる専門的な声を選択できるようにすることです。
2. 開発ガイドライン、選択基準、および声優ランキングに関する初期研究

アフリカ系アメリカ人から倫理上の問題やガイドラインを収集しました。
適切な代表者を選択し、TTSシステムを作成しました。
3. TTSシステム概要

アコースティックモデルとバコーダーからなるTTSアーキテクチャが使用されました。
AA voiceモデルはMS（マルチスピーカー）モデルが選択されました。
4. 合成プロトタイプ生成

AA voiceモデルは他言語版も含めたMSモデルで訓練されました。
MSモデルがSS（単一話者）モデルよりも品質が向上したことが示唆されました。
5. AA VOICEおよびWH VOICEの評価：Study 1 & Study 2
6. アフリカ系アメリカ人向けフォーカスグループ：Study 3

Stats

"米国英語話者はAA TTS音声を正しく識別できませんでした。"
"AA TTS音声サンプル7つ中6つが「White」に関連付けられました。"

Quotes

Key Insights Distilled From

Creating an African American-Sounding TTS

by Claudio Pinh... at arxiv.org 03-19-2024

https://arxiv.org/pdf/2403.11209.pdf

Deeper Inquiries

どうして米国英語話者はAA TTS音声を正しく識別できなかったのか？

Study 1 and Study 2 revealed that U.S. English speakers had difficulty correctly identifying the African American (AA) Text-to-Speech (TTS) voice as being from an African American person. There are several possible reasons for this:

Stereotypical Expectations: Participants may have preconceived notions about how an African American voice should sound, such as expecting certain speech patterns or accents commonly associated with African Americans. When the AA voice did not conform to these stereotypes, participants may have been unable to attribute it to an African American speaker.

Unfamiliarity with Diverse Voices: Due to limited exposure to diverse voices in media and technology, participants may not have been familiar with the range of vocal characteristics present within the African American community. This lack of exposure could lead to difficulties in accurately recognizing a specific racial or ethnic identity based on voice alone.

Implicit Bias: Participants' implicit biases or subconscious associations between race and speech patterns may have influenced their perceptions of the AA voice. If individuals hold stereotypes about how different racial groups speak, they may unconsciously apply these biases when evaluating voices.

Quality of Audio Samples: The quality of the audio samples themselves, including factors like clarity, enunciation, and prosody, could also impact participants' ability to correctly identify the race or ethnicity of the speaker.

Overall, these findings highlight the complex interplay between language perception, cultural stereotypes, and individual biases in determining how people interpret and categorize voices based on race.

WH TTS音声サンプル全てが「White」と関連付けられた理由は何か？

The results showing that all audio samples from the White (WH) Text-to-Speech (TTS) voice were consistently matched with faces representing White individuals can be attributed to several factors:

Cultural Norms: In societies where Whiteness is often considered as a default or standard category for representation in media and technology platforms, individuals may default to associating generic voices with White identities unless explicitly stated otherwise.

Lack of Diversity Awareness: Similar to issues faced by U.S English speakers in identifying the AA TTS voice correctly due to limited exposure to diverse voices; participants might not consider alternative racial identities when matching voices with faces if they are accustomed only to hearing predominantly White representations.

Confirmation Bias: Once participants made an initial association between a particular face and a perceived "White" characteristic for one sample from WH TTS voice; this bias could influence subsequent matches without considering other possibilities objectively.

4 .Ingrained Stereotypes: Deep-rooted societal stereotypes linking certain speech patterns or qualities exclusively with specific races can subconsciously guide individuals towards making assumptions based on superficial cues rather than nuanced understanding.

この研究結果から得られる洞察や今後の展望は？

From these research findings emerge valuable insights into human perceptions related
to AI-generated synthetic voices:

Diversity Representation: The studies underscored challenges around diversity representation in AI technologies,
highlighting gaps in inclusive design practices that need addressing.

Implicit Bias Awareness: Understanding how implicit biases impact our interactions with technology is crucial
for developing more equitable systems moving forward.
-Future Directions: Future research should focus on enhancing diversity awareness among users,
improving algorithms for accurate identification across various demographic categories,
and promoting inclusivity through responsible AI development practices.