MELD: Mel-Spectrogram-Based Speech Language Modeling with Discrete Latent Variables
About
Recent speech language models rely on encoders that are optimized separately from autoregressive models. Since these encoders are unaware of the downstream objectives, the extracted representations may not be optimal for downstream tasks. To address this limitation, we introduce a discrete latent variable model on mel spectrograms that jointly optimizes the encoder and the speech language model. Joint optimization not only brings improvements over codec-based and other mel-spectrogram-based baselines on zero-shot Text-to-Speech (TTS) and Speech-to-Text (STT) tasks, but also effectively alleviates common issues in autoregressive mel-spectrogram modeling, such as prolonged silence generation and word omissions.
Sung-Lin Yeh, Wei Zhou, Gil Keren, Duc Le, Zhong Meng, Hao Tang, Jay Mahadeokar, Ozlem Kalinli, Alexandre Mourachko• 2026
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Text-to-Speech | LibriSpeech clean (test) | WER1.9 | 88 | |
| Automatic Speech Recognition | LibriSpeech 960h (dev-other) | WER9 | 56 | |
| Speech Recognition | LibriSpeech 960 clean (test) | WER3.5 | 24 | |
| Zero-shot Text-to-Speech | LibriSpeech LS960 (test-clean) | WER (Whisper-large)1.9 | 8 | |
| Speech-to-Text | LibriSpeech LS960 (dev clean) | WER3.6 | 6 | |
| Speech-to-Text | LibriSpeech LS960 (test-other) | WER9.2 | 6 | |
| Text-to-Speech | LibriSpeech (test-other) | WER2.4 | 5 | |
| Speech Synthesis | LibriSpeech clean (test) | SMOS3.89 | 4 |
Showing 8 of 8 rows