音訊人工智慧指南

語音助理和口吃

Voice assistants like Siri, Alexa, and Google Assistant are typically trained and tuned to expect fluent, continuous speech, so they often interrupt, mishear, or give up on people who stutter, treating disfluencies like blocks, repetitions, or prolongations as the end of a command or as unrecognized noise.

聆聽此指南AI 語音敘述 · ~3 min

本頁4 分鐘閱讀
  1. 概述
  2. 深入探討
  3. 現實世界的實施
  4. The Future of Voice Assistants and Stuttering
  5. 付諸實踐
  6. 不斷探索
  7. 常見問題

概述

This matters because it affects everyday tasks like setting reminders or making calls, and effectively excludes people who stutter from tools marketed as universally convenient; researchers have specifically built stuttered-speech datasets to help fix this.

深入探討

Stuttering is a speech disorder involving disruptions in the flow of speech, including sound or syllable repetitions, prolonged sounds, and blocks where sound stops entirely for a moment, and it affects a meaningful share of the population at some point, with many people continuing to stutter into adulthood. Speech interfaces combine automatic speech recognition with an endpoint or voice-activity decision about when a person has finished. A pause during a block or a repeated sound may contribute to an early cutoff or a transcription error in some systems, as the cited user study illustrates; the effect varies by person, device, and task. This is a known and documented gap rather than a hypothetical one. Researchers have published work on stuttered speech. SEP-28k is a research dataset with labeled clips of events such as blocks, prolongations, and repetitions; it was created for studying event detection, not as proof that consumer assistants now handle stuttering reliably. A 2023 study surveyed 61 people who stutter and tested speech recognition with recordings from 91 participants, reporting cutoffs and transcription errors in the systems it evaluated. These findings document a gap in those study conditions, not the performance of every current assistant. Related accessibility efforts, such as Google's Project Euphonia, focus more broadly on speech that differs from typical fluent patterns, including stuttering alongside other speech differences. A common misconception is that stuttering is simply 'talking slowly' and that giving an assistant more time to listen fully solves the problem; in reality, disfluencies can include sounds the recognizer misclassifies as noise or as the wrong word entirely, not just longer pauses, so fixing the experience requires models trained on actual disfluent speech patterns, not just adjusted timeout settings. Progress has been described by researchers as incremental, since disfluency patterns vary significantly between individuals.

現實世界的實施

A person who stutters says 'Set a t-t-t-timer for ten minutes,' and an assistant tuned for fluent speech may cut off listening after the first pause, register only part of the phrase, or ask the person to repeat themselves.

Some assistants offer an extended-listening or longer-pause setting that gives more time before deciding the user has finished speaking, which can help but isn't specifically designed for stuttering.

Google has published research and datasets such as the Stuttering Events in Podcasts corpus, using real recorded stuttered speech to train and evaluate models on repetitions, blocks, and prolongations rather than only fluent speech.

Some users who stutter report developing workarounds, like typing commands instead of speaking them, or waiting for a stutter to pass before starting a command, effectively adapting their behavior around a tool that wasn't built with them in mind.

The Future of Voice Assistants and Stuttering

As of September 26, 2026, research continues on more inclusive speech recognition, including datasets and community-informed evaluation. A dataset or laboratory improvement does not guarantee that a particular phone or voice assistant has adopted it. Test the actual device and language, offer typing or touch controls as alternatives, and let users correct a transcript without being forced to repeat. Clinical treatment and communication preferences belong to the individual and a qualified speech-language professional. A failed command should trigger an accessible alternative, not pressure to speak fluently.

付諸實踐

為什麼這很重要

  • 交通與覆蓋範圍. 它透過轉錄、旁白和語音介面提高了可訪問性。
  • 成本與預算. 媒體團隊可以用更少的預算更快地交付精美的音訊。
  • 速度與規模. 面向客戶的系統可以處理更大規模的語音互動。

留意

  • 如果未徵得同意,語音濫用和冒充風險就會增加。
  • 由於口音、方言或嘈雜的環境,準確性可能會下降。
  • 如果沒有明確的標籤,合成音訊可能會被誤認為是真實的語音。

後續步驟

  1. 獲得語音捕獲、克隆和重用的明確同意。
  2. 測試不同揚聲器和背景條件下的品質。
  3. 定義人員必須審查或批准輸出的時間。
  4. 標記合成音訊並保留來源記錄以供問責。

不斷探索

免費電子報

獲取每日 AI 簡報

每個工作日上午提供三則經過驗證的 AI 新聞,以淺顯易懂的英文撰寫。永久免費,無廣告。

每個工作日一封電子郵件。一鍵取消訂閱。我們絕不出售或分享您的地址。

測試一下自己

Take the Voice Assistants and Stuttering quiz

每個答案皆有即時回饋,通過課程後可獲得具備可驗證 ID 的可分享證書。

開始測驗

支持免費的 AI 教育。AI Understanding 是一個 501(c)(3) 非營利組織——絕無廣告,絕無付費牆。 捐款

常見問題

What is Voice Assistants and Stuttering?

Voice assistants like Siri, Alexa, and Google Assistant are typically trained and tuned to expect fluent, continuous speech, so they often interrupt, mishear, or give up on people who stutter, treating disfluencies like blocks, repetitions, or prolongations as the end of a command or as unrecognized noise. This matters because it affects everyday tasks like setting reminders or making calls, and effectively excludes people who stutter from tools marketed as universally convenient; researchers have specifically built stuttered-speech datasets to help fix this.

What can happen when a voice assistant hears a stuttering block, per the guide?

A block's near-silence can trigger the assistant's end-of-speech detection prematurely.

What does SEP-28k refer to, according to the guide?

SEP-28k, Stuttering Events in Podcasts, is a labeled dataset used to train and evaluate models on real disfluent speech.

Why is 'stuttering is just talking slowly' called a misconception in the guide?

The issue isn't just timing; certain sounds get misclassified outright, so longer timeouts alone don't fix recognition.

According to the guide, what is voice activity detection (VAD), as explained in the guide?

VAD determines end-of-speech, typically via a silence threshold, which can misfire on stuttering blocks.

According to the guide, what is Project Euphonia's broader focus, per the guide?

Project Euphonia addresses atypical speech broadly, of which stuttering is one example among others.