JA / EN

Academic research on evaluating and improving speech quality in online meetings

Make sure your voice truly gets through — stress-free online meetings for everyone.
…But what does it mean when online meetings are "hard to hear"?

Motivation

In online meetings, sometimes speakers' voices are buried in noise, hard to hear, or simply difficult to concentrate on.

Many meetings also include power hierarchy: even if a manager or other senior person has poor audio quality and is hard to understand, it can be difficult for others to point this out. Even when the meeting itself appears to proceed without major issues, you may later find that AI transcription or minutes contain many errors, making it time-consuming to confirm what was actually said. Situations like these are commonplace, and they strongly affect how comfortable online meetings feel, how easy they are to follow, how lively the discussion becomes, and ultimately the productivity of the organization.

To check how your own voice sounds to others, you have to record yourself using the meeting tool's audio test feature and listen back. In a busy work environment, however, it is hard to make time for this before every meeting, and the judgment of "easy to hear or not" relies entirely on personal, subjective impression. In addition, even if you feel the audio is hard to hear, it is not easy for a typical user to identify the underlying cause.

We believe that, if there were a service that could automatically analyze your audio quality and suggest improvements based on the specific causes of degradation, online meetings could become more comfortable and efficient. Likewise, in AI transcription and automated minutes services, if we could know in advance whether the audio quality is suitable for accurate recognition, we could further improve the productivity of knowledge work.

Problem statement

"If AI can correctly recognize a speech signal, that speech is also easy for humans to understand."

In this research, we start from the hypothesis above. Based on this assumption, we analyze how various factors that degrade audio affect the accuracy of AI transcription. We then train AI models to learn the relationship between each degradation factor and transcription accuracy, exploring the possibility of objectively evaluating the quality of online meeting audio.

Examples of degraded audio

Speech degraded by 40% packet loss
Speech degraded by 40% packet loss
Clean speech (no packet loss)
Clean speech (no packet loss)

Are the tendencies of mishearing different between humans and AI?

Research methodology

A. Machine learning

We use simulation to intentionally generate degraded speech and train AI models to learn the patterns of degradation. We then apply the learned models to real meeting audio to automatically analyze the causes and severity of audio degradation and propose specific improvement strategies based on those findings.

B. Subjective evaluation

This research project is based on the hypothesis: "If AI can correctly recognize a speech signal, that speech is also easy for humans to understand". To verify this, we designed a subjective listening experiment that observes how human listeners actually hear degraded speech.

Among the many factors that affect speech quality, our first step focuses on packet loss in IP networks. We simulated packet loss in Japanese speech signals and conducted a transcription task using these degraded audio samples. By analyzing what participants wrote down and how they did so, we examine how packet loss affects human perception (clarity, intelligibility) and human behavior (number of replays, time required to complete the transcription, etc.).

We then compare the human transcription results with those of an automatic speech recognition (ASR) system and analyze how error patterns differ depending on acoustic and linguistic factors. This allows us to quantitatively understand the differences between how humans and AI listen, and to apply those findings to more advanced automatic speech quality assessment and to human-centered voice interface designs.

Overview
Overview of research methodology

Conferences & publications

Dec 1–5, 2025

Evaluating speech quality for automatic transcription in videoconferencing

6th Joint Meeting Acoustical Society of America and Acoustical Society of Japan

Honolulu, Hawaii, USA
Dec 2025

How Online Meetings Change Your Voice—and How We Measure It

Acoustics Lay Language Papers (ASA)

↗ acoustics.org
Press release
Mar 17–19, 2026

The Effects of Packet Loss in IP Communication on the Transcription Accuracy of Japanese Speech

Acoustical Society of Japan, 2026 Spring Meeting

↗ Poster PDF (Japanese)
Tokyo, Japan · Poster presentation (first day: 3/17)
Sep 8–10, 2026

Comparison of Human and Automatic Speech Recognition Responses to Packet-Loss-Degraded Speech

Acoustical Society of Japan, 2026 Autumn Meeting

Kanazawa city, Ishikawa pref., Japan · Poster presentation (second day: 9/9)
Sep 8–10, 2026

Automatic Speech Recognition–Based Speech Quality Assessment

Acoustical Society of Japan, 2026 Autumn Meeting

Kanazawa city, Ishikawa pref., Japan · Poster presentation (third day: 9/10)