A.1    Text-to-Audio Generation

Training Dateset Details. The training set for Tango-Full includes AudioCaps, WavCaps, UrbanSound, Musical Instrument, MusicCaps, GTZAN, and ESC. All audio clips longer than 10 seconds are segmented into 10-second chunks and resampled to 16kHz. The training set for AudioLDM-Full further includes Adobe Foley, Clotho, FreeSound50K, TUT, DCASE2023 Task 7. We use GPT-4 to convert raw textual descriptions into full captions for uncaptioned samples. In addition, we apply stricter data cleaning procedures: for long recordings, only the first 30 seconds are retained, and samples are filtered using CLAP, retaining only those with a similarity score above 0.4.

A.2    Temporally Controllable Text-to-Audio Generation

Fusion Rate Ablation in CRD. As shown in the Tab. 6, when the fusion step is set to 1.0, temporal controllability (measured by S-CLAP and S-CIDEr) and semantic alignment (measured by CLAP) are lower compared to fusion steps of 0.75 or 0.5. This is likely due to semantic suppression in overlapping regions. As the fusion step decreases, such suppression is reduced, leading to improvements in both temporal and overall performance. However, when the fusion step becomes too low, the number of denoising steps involving region-wise cross-attention is insufficient. This results in content blending within overlapping regions and a lack of global stylistic coherence, ultimately degrading audio quality (FAD) and semantic alignment.

Table 6: Ablation Study on Fusion Rate r in CRD

Qualitative comparison. As shown in Fig. 3, we present qualitative comparison examples of temporally controllable text-to-audio generation methods. As previously discussed, training-free approaches, including DC, CRD, and MD, demonstrate significantly better temporal controllability than Make2, since Make2 only receives coarse subprompt position cues in the form of pairs like <A cat meows \& start> or <A cat meows \& mid>. Moreover, Make2 also struggles with controlling the number of sound occurrences over time, resulting in mismatches between prompt timing and actual sound placement. As shown in Fig. 3 line 4, the A cat meows event occurs only once, and its timing does not align with the reference. Among the training-free methods, MD produces strong noise in overlapping regions, with low fidelity and little discernible content. The boxed overlap regions in Fig. 3 line 3 display that spectrogram is dominated by random noise patterns. In contrast, CRD better preserves clarity, but sometimes shows semantic suppression in overlapping segments. For instance, in the example of Fig. 3 line 2 shown, the A cat meows sound dominates the overlap region, suppressing the presence of the crickets chirp. As a result, the crickets chirp sound appears fragmented and inconsistent in the overlap regions. DC, Fig. 3 line 1, which simply concatenates independently generated regions, avoids this issue due to the complete isolation of subprompts during generation. In the overlap regions, spectrograms of both sound events are clearly distinguishable.

Figure 3: Qualitative comparison result of temporally controllable Text-to-Audio methods. Boxed regions indicate the occurrence of the event A cat meows. Box: correct timing and audio content; Box: semantic suppression; Box: noise; Box: temporal misalignment.

However, as illustrated in Fig. 4 line 2, DC lacks a mechanism to enforce global stylistic consistency, resulting in noticeable variations in timbre, background texture, and unnatural transitions between adjacent regions. The spectrograms of adjacent regions exhibit clear discontinuities. In comparison, CRD, Fig. 4 line 1, maintains a unified global style, with smooth spectrogram transitions across regions and a more coherent auditory experience.

Figure 4: Qualitative comparison result of CRD and DC. Box: stylistic inconsistency or unnatural transitions between regions. Event1: Cheers and applause from the crowd; Event2: Booing and shouting from the crowd, along with scattered applause; Event3: A few scattered shouts from men.