Background: Listeners typically show a 10-15% improvement in speech-in-speech intelligibility when target speech is spoken by someone familiar, compared with someone unfamiliar—both for naturally familiar and for lab-trained voices. This improvement could have real-world benefits, if listeners were trained on voices they would later encounter in a noisy natural setting (e.g., public transport). However, this application depends on intelligibility improvements enduring over a prolonged period, which has not been tested systematically. This study examined whether, after one session of online training, participants showed a speech-in-speech intelligibility benefit for trained voices one day, one week, and four weeks after the initial training session.

Methods: 40 participants were trained to recognise three novel voices. During training, participants listened to each voice produce 166 natural, meaningful sentences (approximately 70 minutes in total). A speech-in-speech intelligibility test was completed immediately after training, one day later, seven days later, and 28 days later. On each trial of the intelligibility test, two voices simultaneously spoke different closed-set sentences, at two possible target-to-masker ratios (+3 and -6 dB TMR). Participants reported the words spoken by the target voice, which was either one of the three trained voices, or an unfamiliar voice. The exact sentences and unfamiliar voices differed across the four test sessions. All tasks were completed remotely online.

Results: Immediately post-training, participants showed an intelligibility benefit of ~13% for trained voices over unfamiliar voices. Critically, this familiar-voice benefit was significant at all timepoints, and the size of the benefit did not significantly change over time. Furthermore, these effects were present at both target-to-masker ratios. Overall intelligibility performance did not differ between the test sessions.

Discussion: Our results show that a single training session produces a substantial speech-in-speech intelligibility benefit for learned voices, which persists for at least four weeks without diminishing. Furthermore, this long-lasting benefit emerges after only 70 minutes of online training on three novel voices and is robust across varying levels of masking speech. These findings indicate that representations of new voices are formed rapidly yet remain stable over time. Our results also suggest that voice familiarisation training could be an effective practical tool for improving intelligibility, as a particular speaker’s voice could be learned several weeks before it is encountered in the real world. In summary, this study demonstrates that brief voice training can produce long-term and practically meaningful improvements in speech-in-noise perception.