Benchmarking Language Modeling for Lossless Compression of Full-Fidelity Audio

Researchers introduce Trilobyte, a byte-level tokenization approach that enables language models to perform lossless audio compression on full-fidelity 16/24-bit audio files. While LMs outperform FLAC at 8 and 16-bit depths, compression gains diminish at higher bit depths, suggesting practical limitations for real-world audio applications.

This research addresses a fundamental scaling problem in applying language models to audio compression. Previous work demonstrated that autoregressive models could compress 8-bit audio effectively, but the vocabulary explosion at higher bit depths—reaching 16.7 million tokens for 24-bit audio—made the approach computationally infeasible. Trilobyte's byte-level tokenization elegantly solves this by reducing vocabulary scaling from exponential O(2^b) to constant O(1), unlocking the first tractable 24-bit LM-based compression system.

The work builds on the broader trend of repurposing transformer-based language models for domain-specific tasks beyond natural language. As computational efficiency improves and model architectures mature, researchers increasingly apply these tools to sequential data across modalities. This reflects growing confidence that the underlying principles of next-token prediction transfer across domains.

For the audio codec industry, the results present mixed signals. LM-based compression decisively beats FLAC at 8-bit and 16-bit resolutions, which could motivate adoption in specialized applications like speech processing or streaming services. However, the diminishing returns at 24-bit—where compression gains plateau—suggests LMs may not displace established codecs like FLAC or specialized algorithms for professional audio workflows. The computational overhead of running inference through large language models also poses practical deployment challenges compared to lightweight traditional codecs.

Future research should focus on optimizing inference speed, exploring hybrid approaches combining LMs with traditional compression, and testing performance on extremely high-fidelity formats. The findings highlight that architectural innovations alone cannot overcome fundamental information-theoretic constraints at the highest bit depths.

• →Trilobyte byte-level tokenization reduces vocabulary scaling complexity from exponential to constant, enabling first-ever 24-bit LM-based lossless audio compression.
• →Language models outperform FLAC at 8-bit and 16-bit audio depths but show diminishing compression gains as bit depth increases beyond 16-bit.
• →The approach addresses a critical scaling limitation that previously restricted LM-based audio compression to 8-bit audio only.
• →Practical deployment remains challenged by computational overhead compared to lightweight traditional codecs like FLAC.
• →Results suggest LMs excel in lower-fidelity audio compression but may not displace established codecs for professional high-fidelity workflows.