Why Audio Language Models Hallucinate Words in TTS

Modern text-to-speech (TTS) systems increasingly rely on autoregressive audio language models, which occasionally generate words, filler sounds, or entire sentences that never existed in the input prompt. This phenomenon, known as semantic hallucination, undermines the fidelity of speech synthesis. This article examines the core technical causes behind these phantom vocalizations, focusing on autoregressive decoding mechanics, cross-attention drift, training data contamination, and stochastic sampling parameters.

The Shift to Generative Autoregressive Architectures

Traditional TTS architectures relied on deterministic alignments, such as duration predictors in non-autoregressive acoustic models (e.g., FastSpeech). These systems mapped phonemes directly to spectrogram frames, making it structurally impossible to generate extra words.

Modern systems, such as VALL-E, Bark, and other discrete-token audio language models, frame speech synthesis as a conditional sequence-to-sequence language modeling task. In this paradigm, acoustic tokens are generated sequentially based on probabilities. Because the architecture behaves like a large language model (LLM), it is inherently generative rather than strictly deterministic. If the predictive confidence of the text conditioning weakens during generation, the model's internal language prior takes over, predicting what sounds natural to say next rather than adhering to the source text.

Predictive Completion and Linguistic Prior

Audio language models are trained on massive text and audio corpora, absorbing natural linguistic flow and conversational conventions. When an input text prompt feels pragmatically incomplete or structurally conversational, the underlying model may attempt to "complete the thought."

For example, if an input sentence ends abruptly, the model may hallucinate closing phrases like "thank you," "goodbye," or affirmative acknowledgments ("yeah," "okay"). The model's language prior overrides the text conditioning constraint because the statistical probability of a conversational sign-off following a spoken utterance is high in the training distribution.

Cross-Attention Breakdown and Alignment Drift

Most audio language models use cross-attention layers to condition the generation of audio tokens on input text tokens. Semantic hallucinations frequently occur when this attention mechanism loses tracking:

• Attention Slippage: As the output sequence grows longer, the attention weights between the generated audio tokens and the input text tokens can disperse or collapse.
• Loss of Monotonicity: Speech naturally progresses monotonically through text. Standard multi-head attention does not strictly enforce monotonic alignment. If the model fails to attend to the appropriate remaining text tokens, the decoder continues to emit speech tokens based entirely on its own previous audio outputs, generating ungrounded semantic content.

Training Data Imperfections and Label Noise

Audio language models require thousands of hours of speech paired with transcripts. These datasets are often scraped from audiobooks, podcasts, and video platforms, with text generated via automated speech recognition (ASR) or loosely matched subtitles.

This introduces label noise:

• Unlabeled Audio: If the training audio contains conversational fillers ("um," "uh"), throat clearing, or trailing sentences that were omitted from the paired transcript, the model learns that producing spoken words from "silent" or non-existent text tokens is acceptable.
• Mismatched Endings: Cutoffs where audio continues past the final transcribed word condition the model to hallucinate trailing audio at the end of synthesis runs.

Exposure Bias and Autoregressive Sampling

During training, models typically use teacher forcing, where each step receives the ground-truth previous token. During inference, however, the model conditions on its own previously predicted tokens.

If the model produces a slight phonetic error or an unintended micro-pause, this unexpected token is fed back into the context window. This divergence from the training distribution—known as exposure bias—can snowball. The model enters an out-of-distribution state where its text-conditioning breaks down, leading the decoder into a self-reinforcing loop of generating coherent-sounding, hallucinated speech.

Furthermore, inference sampling strategies like high temperature, Top-K, or Top-P (nucleus) sampling introduce randomness to make speech sound more expressive and less robotic. Higher entropy during decoding increases the statistical likelihood of sampling unexpected semantic tokens, turning acoustic variance into hallucinated words.