Literature searches were performed across five Chinese and English databases, namely PubMed, Cochrane Library, Web of Science, Embase, and CNKI. Owing to the insufficient number of studies, we expanded the search by using a combination of subject terms and free words. The search terms were as follows: “Speech Disorders or anomia* or aphasi*,” “music therapy or melodic intonation therapy or rhythmic speech cueing or vocal intonation therapy or sing*.” We conducted searches from database inception to July 2023, and further supplemented by reviewing the references and grey literature cited in the included studies. The details of the search strategy is described in Supplementary material 1.

Our inclusion criteria were as follows: (1) P: Adults aged 18 years and older with non-fluent aphasia; (2) I: NMT (e.g., MIT, rhythmic speech cueing, musical speech stimulation, therapeutic singing, and voice pitch therapy); (3) C: Other speech training or no training; and (4) O: Speech functions including naming, comprehension, repetition, spontaneous speech, and communication. We included RCT studies that followed the PICO format and were written in English or Chinese. Studies that combined music therapy with other interventions (such as drug therapy, artificial intelligence, transcranial magnetic stimulation, and transcranial direct current stimulation), those that solely involved listening to music as an intervention, those without outcome indicators related to speech functions, those without a full text and conference papers, were excluded.

Four authors (JYG, WL, SQZ, CJL) were involved in screening the studies. Duplicates were simultaneously removed by the two authors (WL, CJL), and literature screening was independently performed by the other two authors (JYG, SQZ). Irrelevant studies were excluded after reading the study titles and abstracts, followed by full-text reading and review of potential studies for inclusion. Inconsistencies in study inclusion results were resolved by another author (XYZ).

Information, including country of publication, design type, sample size, disease type, aphasia type, intervention method, intervention dose, and outcome measures, were included in the table of study characteristics.

The quality of the included studies and risk of bias for each study were assessed according to the Cochrane Collaboration’s recommendations (18), which included selection, performance, detection, attrition, and reporting bias. Quality control and bias assessment were independently performed by the two authors (JYG, SQZ). In case of disagreements, consensus was reached through a discussion (Supplementary material 2).

Literature reading was independently performed by the two authors (JYG, SQZ) and data related to the study outcome indicators were extracted (Supplementary material 2). Another author (WL) checked and imported the data into the Review Manager V5.4.1 (Copenhagen: The Nordic Cochrane Centre, The Cochrane Collaboration, 2012) for meta-analysis.

Five outcome measures namely, naming, comprehension, repetition, spontaneous speech, and communication, were included in the meta-analysis. The assessment scales for the outcome measures included the Boston Diagnostic Aphasia Examination (BDAE), Amsterdam Nijmegen Everyday Language Test (ANELT), Western Aphasia Battery (WAB), Communicative Activity Log (CAL), and Chinese Rehabilitation Research Center Standard Aphasia Examinations (CRRCAE).

Statistical analyses were performed on the Review Manager V5.4.1 (RevMan V5.4.1) and Stata/MP 17.0. As different assessment tools were used; the effect size of measurement data was represented using the standardized mean difference (SMD). The SMD and 95% confidence interval (CI) of the pre- and post-intervention differences were employed to express the continuous variable outcomes. All data were subjected to pool analysis using a random-effects model. The heterogeneity of the included studies was tested based on the I² values. An I² value >50% and p < 0.1 indicated a high level of heterogeneity, and further subgroup analysis was required to determine the source of heterogeneity. Sensitivity analysis was performed on each subgroup using the leave-one-out method. The significance level for meta-analysis was α = 0.05. Publication bias was analyzed using funnel plot. A forest plot analysis was performed by pooling the same outcome indicators, and sensitivity analysis was performed on the outcome measures for each group.

A total of 8,635 studies were retrieved, and the studies were supplemented by checking the citations. After checking for duplicates, reading the abstracts, and subsequently the full text for precise screening, a total of 11 studies were finally included. The literature screening process is shown in Figure 1. The list of studies that were excluded after refined screening is provided in Supplementary material 3.

The eleven studies (14–17, 19–25) comprise 329 patients, with 167 and 162 patients in the experimental (EG) and control group (CG), respectively. The between-group comparisons of general information across all included studies did not reveal any significant differences. Regarding outcome measures, seven studies assessed the patients’ naming, nine assessed their comprehension, ten assessed their repetition, five assessed their spontaneous speech, and three assessed their communication. The basic characteristics of the included studies are provide in Table 1.

The quality of the 11 included studies was assessed using the Cochrane Handbook (26) and the results are shown in Figure 2. All studies provided clear description of the generation of random sequences. Three studies (21, 23, 24) explicitly described the methods for allocation concealment. All studies explicitly reported comparability at baseline between groups. Performance bias was marked as uncertain in most studies due to the specificity of the intervention, making it impossible to completely blind the investigators and participants. Most studies followed the design principle of blinding the assessors. All studies presented complete outcome data, with no selective reporting.

Six studies (15, 21–25) assessed patients’ naming ability. A meta-analysis using a random-effects model showed that the difference between the two groups was not statistically significant (SMD = 0.28, 95%CI [−0.04, 0.60], p > 0.05), suggesting that the EG did not differ from the CG with respect to enhancing patients’ naming ability (Figure 3).

Eight studies (14–16, 19, 20, 22, 23, 25) assessed the patients’ comprehension. Meta-analysis using a random-effects model demonstrated that the difference between the two groups was not statistically significant (SMD = 0.09, 95%CI [−0.16, 0.35], p > 0.05), indicating that the EG did not differ from CG with respect to enhancing patients’ comprehension (Figure 4).

Nine studies (14, 15, 17, 19, 20, 22–25) assessed patients’ repetition. Meta-analysis using a random-effects model showed that the difference between the two groups was statistically significant (SMD = 0.37, 95%CI [0.12, 0.62], p < 0.05), indicating that the EG was superior to CG in enhancing patients’ repetition (Figure 5).

Four studies (16, 21, 22, 25) assessed patients’ spontaneous speech. Meta-analysis using a random-effects model showed that the difference between the two groups was not statistically significant (SMD = 0.30, 95%CI [−0.03, 0.62], p > 0.05), suggesting that the EG did not differ from the CG with respect to enhancing the patients’ spontaneous speech (Figure 6).

Three studies (15, 16, 20) assessed patients’ communication. A meta-analysis using a random-effects model showed that the difference between the two groups was not statistically significant (SMD = 0.34, 95%CI [−0.09, 0.77], p > 0.05), suggesting that the EG did not differ from CG with respect to enhancing patients’ communication (Figure 7).

Exploratory grouping of intervention doses was performed on the included studies that examined repetition (Supplementary material 4). Meta-integration of any study with an intervention dose of 30 h and ≤ 20 h was performed using a random-effects model, which all resulted in statistically significant differences between the music therapy group and other intervention methods. Meta-integration using a random-effects model of studies with intervention doses ≤30 h but not ≤20 h, yielded statistically significant differences between the music therapy group and other intervention methods (Supplementary material 4). Therefore, an intervention dose of 20 h was selected as the cut-off for subgroup analysis to compare the effect of NMT on the repetition of patients receiving different intervention doses, and a meta-analysis was performed using a random-effects model. Five studies (17, 19, 20, 23, 25) with NMT intervention doses ≤20ours h showed no significant difference between the two groups (SMD = 0.32, 95%CI [−0.01, 0.66], p > 0.05), indicating that the EG did not differ from the CG with respect to enhancing patients’ repetition. Four studies (14, 15, 22, 24) with NMT intervention doses >20 h showed a statistically significant difference between the two groups (SMD = 0.43, 95%CI [0.06, 0.79], p < 0.05), indicating that the EG was superior to the CG in enhancing patients’ repetition (Figure 8).

The I² of all studies was >50% (p > 0.05), indicating no significant heterogeneity among the pooled studies and no further subgroup analysis was required. Stata/MP 17.0 for Windows (Copenhagen: Computer Resource Center, 2021) was used to conduct sensitivity analysis on the outcome measures using the leave-one-out method. The results showed that for repetition, after any one study was excluded, the pooled results of the remaining studies remained statistically significant (95%CI not including 0). This was consistent with the original pooled results, indicating that the results were stable. For naming, comprehension, and communication, after any one study was excluded, the pooled results of the remaining studies remained statistically insignificant (95%CI including 0). This finding was consistent with the original pooled results, indicating that the results were stable. For spontaneous speech, after any one of three out of four included studies was excluded, the pooled results of the remaining studies were not statistically significant (95% CI including 0), which was consistent with the original pooled results, indicating that the results were stable (Supplementary material 5).

Using the rating method recommended by the GRADE system (27), the quality of evidence was assessed based on five domains: risk of bias, inconsistency, indirectness, imprecision, and publication bias. The results showed that the methodological quality of the included literature was moderate for naming, comprehension, repetition, and communication, and high for spontaneous speech (Table 2). The Cochrane Risk of Bias Assessment Tool was used to assess the limitations of the included RCTs. When the total weight ratio of high-risk literature was >20% due to irrational design or unclear descriptions of randomization, allocation concealment, blinding (e.g., implementation and measurement), then downgrading should be considered. Therefore, naming, comprehension, and repetition were each downgraded by one level for risk of bias. The pooled I² for all outcomes was >50% (p > 0.05), and the differences in the intervention types, doses, and durations among the included studies were small; therefore, no study results were downgraded for heterogeneity. The patient populations, interventions, and important outcome indicators were largely consistent across all included RCTs, and hence no study results were downgraded for indirectness. In terms of imprecision, the quality of evidence for communication was downgraded due to wide CI or limited sample size. Publication bias was visually evaluated using funnel plot. Repetition, which had the largest number of included studies, was selected as the main outcome to create the funnel plot. Our findings revealed that most studies had large sample sizes and were concentrated in the narrow, upper part of the funnel plot, suggesting that the result was relatively robust. Furthermore, Egger’s test yielded p = 0.251, and thus the result was not downgraded. The funnel plot is shown in Figure 9.