SonoEdit: Null-Space Constrained Knowledge Editing for Pronunciation Correction in LLM-Based TTS

TL;DR

SonoEdit tackles mispronunciation in LLM-based TTS by surgically editing pronunciation representations without retraining. It first localizes pronunciation-related information to mid-to-late transformer layers using Acoustic Causal Tracing, then applies a null-space constrained, closed-form weight update to correct target pronunciations while preserving general speech in orthogonal subspaces. The method requires no extra parameters and operates in one shot, yielding strong pronunciation improvements (e.g., on HardNoun-300) with minimal disruption to prosody, speaker identity, and overall WER. This approach offers a practical, data-efficient pathway for deploying linguistically diverse TTS systems, while highlighting constraints around null-space generalization and the need for scalable editing across batches.

Abstract

Neural text-to-speech (TTS) systems systematically mispronounce low-resource proper nouns, particularly non-English names, brands, and geographic locations, due to their underrepresentation in predominantly English training corpora. Existing solutions typically rely on expensive multilingual data collection, supervised finetuning, or manual phonetic annotation, which limits the deployment of TTS systems in linguistically diverse settings. We introduce SonoEdit, a model editing technique that surgically corrects pronunciation errors in pre-trained TTS models without retraining. Instead of costly finetuning or explicit phoneme injection, we propose a parsimonious alternative based on Null-Space Pronunciation Editing, which performs a single-shot parameter update to modify the pronunciation of specific words while provably preserving all other model behavior. We first adapt Acoustic Causal Tracing to identify the Transformer layers responsible for text-to-pronunciation mapping. We then apply Null-Space Constrained Editing to compute a closed-form weight update that corrects the target pronunciation while remaining mathematically orthogonal to the subspace governing general speech generation. This constrained update steers the model's acoustic output toward a desired pronunciation exemplar while guaranteeing zero first-order change on a preserved speech corpus.

Figures (4)

• Figure 1: A schematic diagram illustrating an LLM-based speech generation pipeline. A Raw Waveform is first processed by a Tokenizer to produce a sequence of Discrete Tokens. These tokens are then input into a Large Language Model (LLM), which generates a sequence of New Tokens. Finally, a Decoder converts the new tokens into a Speech Waveform.
• Figure 2: An illustration of null-space constrained editing. The pronunciation correction is applied in directions orthogonal to the subspace representing general speech characteristics, ensuring invariance of non-target behavior.
• Figure 3: Three complementary analysis methods show convergent evidence that layers 15-21 encode phonetic information: causal mediation analysis (Indirect Effect, red), linear probe classification (Probe Accuracy, blue), and gradient-based attribution (Gradient Norm, green).
• Figure 4: Qualitative comparison of generated audio samples.