TL;DR
This paper systematically investigates how the choice of audio segment length affects the performance of neural audio fingerprinting systems. The authors reveal that shorter segments (0.5 seconds) consistently outperform longer ones, particularly for short query durations, and demonstrate that GPT-5-mini can reliably recommend these optimal configurations compared to other leading LLMs.
What Problem Does This Paper Address?
Audio fingerprinting is the technology behind applications like Shazam, creating compact, unique digital signatures of acoustic signals for rapid identification and retrieval. To capture local acoustic features, deep learning models typically slice continuous audio into short, fixed-duration segments.
However, the choice of this segment length ($W$) has traditionally been treated as a heuristic afterthought rather than a controlled variable. Does a 1-second chunk contain the optimal balance of temporal resolution and acoustic context, or would shorter or longer windows yield better retrieval accuracy? This paper fills this gap by analyzing the interplay between segment length, query duration, and retrieval performance.
Methods
The authors build upon a state-of-the-art baseline called the Neural Audio FingerPrinter (NAFP), which natively operates on fixed 1-second segments. To evaluate different segment lengths ($W \in \{0.5, 1, 2\}$ seconds) without completely altering the underlying architecture, they introduce NAFP+.
NAFP+ inserts a fully-connected layer with an Exponential Linear Unit (ELU) activation function right before the model's primary convolutional blocks. This addition acts as a normalization step, mapping the variable number of Short-Time Fourier Transform (STFT) frames ($T$) resulting from different segment lengths back to a fixed dimension ($T_0 = 32$).
During the query phase, a query of length $L$ is broken down into $W$-second segments using a sliding window with a 0.5-second hop size. The system retrieves the closest matches from the database for each segment and uses a weighted majority voting scheme to determine the final correct audio track.
Datasets
The experiments were conducted using a music dataset derived from the Free Music Archive (FMA) collection.
- Training Set: 10,000 audio clips (83.3 hours).
- Reference Set ($D_{ref}$): 500 audio clips (4.2 hours) used to generate queries.
- Distractor Set ($D_{dist}$): 9,978 audio clips (83.3 hours) acting as noise to ensure the retrieval task is non-trivial.
Queries were realistically corrupted by overlaying background noise, introducing time offsets, and simulating room impulse responses (reverb).
Key Insights and Ideas
- Shorter is Better for Precision: Slicing audio into 0.5-second segments provides superior temporal resolution. This granularity is particularly potent when the user's query is very short ($\le 3$ seconds).
- Context Compensation: As the query length grows longer than 3 seconds, the accuracy gap between 0.5-second and 1-second segments narrows significantly. This suggests that a longer total query duration provides enough macro-level temporal context to compensate for coarser individual segments.
- LLMs as Architecture Advisors: The authors also explored whether Large Language Models could predict the optimal segment length through zero-shot prompting. Interestingly, different models showed vastly different capabilities in grounding their design advice in empirical reality.
Key Results
The empirical results clearly favor shorter segment windows across multiple metrics, including Top-K Exact Hit Rate and Top-1 Near Hit Rate:
- Dominance of 0.5s Windows: The 0.5-second NAFP+ model achieved the highest win rate, securing the top spot in up to 10 out of 10 evaluated query lengths for the Top-10 Exact Hit Rate metric.
- Diminishing Returns on Query Length: Across all segment lengths, retrieval performance improves rapidly with longer queries but begins to saturate after 4 seconds.
- LLM Evaluation: When tested across 5 distinct prompts evaluating different engineering trade-offs (e.g., accuracy vs. storage), GPT-5-mini consistently recommended a 1-2 second sweet spot, aligning closely with the empirical data. In contrast, Gemini-2.5-flash and Claude-Sonnet-4.5 tended to recommend longer, less performant segments (up to 8 seconds).
Limitations and Future Work
While the paper provides clear guidance for music retrieval, the authors note that these findings might not immediately generalize to other audio domains like speech processing or environmental sound identification.
Future research directions include testing these segment dynamics on tasks like speaker verification and keyword spotting, as well as developing adaptive, lightweight predictors that can dynamically alter segment lengths based on the real-time content or length of the incoming query.
Summary updated