Evaluating synthesized speech intelligibility in noise

Area of Science:

• Speech processing
• Human-computer interaction
• Acoustic phonetics

Background:

• Humans adapt speech for clarity in noise.
• Advancing speech synthesis technology enables machines to generate intelligible speech.
• Evaluating synthesized speech intelligibility in noise is crucial for human-computer interaction.

Purpose of the Study:

• To assess the subjective and objective intelligibility of synthesized speech in speech-shaped noise.
• To compare the intelligibility of synthesized voices from major platforms against human speech.
• To evaluate the performance of automatic speech recognition systems with synthesized speech.

Main Methods:

• Subjective intelligibility tests involving human listeners.
• Objective intelligibility measurements.
• Evaluation of synthesized speech from three major speech synthesis platforms.
• Testing with speech-shaped noise.

Main Results:

• Synthesized voices demonstrated an intelligibility range comparable to human voices.
• Certain synthesized voices exhibited higher intelligibility than human voices.
• Two modern automatic speech recognition systems achieved 10% higher word recognition accuracy compared to human listeners.

Conclusions:

• Speech synthesis technology can produce voices highly intelligible in noisy environments.
• Synthesized voices show potential to match or exceed human speech intelligibility.
• Automatic speech recognition systems perform exceptionally well with intelligible synthesized speech.