Source-linked AI summary
SonoEdit: Null-Space Constrained Knowledge Editing for Pronunciation Correction in LLM-Based TTS
Ayush Pratap Singh, Harshit Singh, Nityanand Mathur, Akshat Mandloi, Sudarshan Kamath
TL;DR
Rare proper nouns are often mispronounced by TTS systems, limiting applications that require accurate pronunciation. SonoEdit localizes pronunciation processing with causal tracing and applies a closed-form null-space edit; on HardNoun-300, it reports a 91% correction rate with negligible impact on global word error rate and speaker identity.
Problem
TTS systems struggle to pronounce proper nouns and other words outside their training distribution, while existing correction approaches can require manual annotation, task-specific training, or unreliable prompting.
Source: §1 · p. 1, §2.1 · p. 2Method
SonoEdit uses Acoustic Causal Tracing to locate pronunciation-related Transformer layers and applies a closed-form, null-space-constrained update without retraining.
Source: §1 · pp. 1–2, §3.1 · p. 4, §3.2 · p. 5Results
91% correction rate on HardNoun-300, with Global-WER at 3.15% and speaker identity SIM at 0.99, while significantly outperforming unconstrained methods.
Source: §5 · p. 9Takeaways & Limitations
SonoEdit offers a one-shot approach for correcting rare-entity pronunciations while preserving general speech performance and speaker identity in the reported experiments.
Source: §5 · p. 9SonoEdit relies on null spaces precomputed from a representative speech corpus, and deployment-distribution shifts could weaken its orthogonality guarantees.
Source: §5 · p. 9
Abstract
from arXiv · showhide
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.
1. Introduction
SonoEdit asks where pronunciation is represented in LLM-based TTS and how to edit specific texts without changing other model behavior. It combines layer localization with null-space-constrained, one-shot updates.
1. Introduction
- TTS systems often mispronounce proper nouns, technical terms, and neologisms outside their training distribution, limiting deployment where accurate pronunciation matters.Existing approaches include manual phoneme annotation, explicit phoneme inputs, fine-tuning, PEFT, and prompt engineering, each with stated costs or shortcomings.Source: §1 · p. 1
- SonoEdit addresses how to locate pronunciation in an LLM-based TTS system and surgically modify specific texts while preserving other behavior.The introduction presents this as its central research question.Source: §1 · p. 1
- Acoustic Causal Tracing identifies Transformer layers responsible for grapheme-to-phoneme mappings, allowing more focused corrections than layer-wise or module-level interventions.The identified layers can be a narrow range, such as layers i to j.Source: §1 · pp. 1–2
- Null-Space Constrained Editing computes a closed-form update orthogonal to general-speech features, while adding no trainable parameters and requiring no training loop.The introduction says this constraint prevents effects on behavior beyond the target pronunciation and describes the update as one-shot.Source: §1 · p. 2
2. Related Works
Related work describes LLM-based TTS and pronunciation-editing approaches, then positions SonoEdit as a transfer of knowledge-editing ideas that avoids PEFT’s training and preservation limitations.
2.1. LLM-based Text-to-Speech Modelling
- LLM-based TTS uses discrete audio tokens and captures linguistic dependencies for nuanced prosody, yet still mispronounces rare, out-of-distribution proper nouns.Such errors matter in navigation, news reading, and multilingual voice assistants, where accurate named-entity pronunciation supports comprehension.Source: §2.1 · p. 2
2.2. Locating and Editing Word-Pronunciation Associations in TTS
- Prior TTS work probes representations and intervenes in layers, while SonoEdit traces G2P layers and constrains edits to be orthogonal to general-speech characteristics.Crude perturbations and layer-wise interventions can degrade general acoustic capability or speaker identity.Source: §2.2 · pp. 2–3
2.3. Knowledge Editing
- Knowledge-editing methods such as ROME [17], MEMIT [18], MEND [19], SERAC [19], and AlphaEdit [20] target factual associations while seeking to preserve general capabilities.SonoEdit adapts these ideas to TTS by applying causal tracing and null-space editing to pronunciation mappings.Source: §2.3 · p. 3
2.4. Parameter Efficient Fine-tuning
- LoRA [21] and related PEFT methods reduce trainable parameters, but pronunciation correction can require target-specific data and still risk acoustic degradation.SonoEdit instead edits in the null-space of general speech, requiring neither retraining nor additional parameters.Source: §2.4 · pp. 3–4
3. Methodology
SonoEdit treats pronunciation errors as faulty internal associations and edits the coarse-token planning pathway rather than acoustic components. It localizes pronunciation-sensitive layers, then constrains target corrections to directions that minimally affect general speech.
3.1. Preliminaries
- Because pronunciation planning resides in coarse tokens, SonoEdit treats errors as faulty hidden-state associations and edits LLM representations, not acoustic components.Each frame also contains fine acoustic tokens, but the method restricts causal analysis and editing to coarse tokens.Source: §3.1 · p. 4
3.2. Acoustic Causal Tracing
- Causal tracing corrupts input representations and restores clean layer activations, locating pronunciation-sensitive layers typically in a contiguous block from L/2 to 3L/4.The layer impact is defined by recovery in the probability of the correct coarse token when clean activations are restored.Source: §3.2 · p. 5
3.3. Null Space: Importance and Utilization
- The method derives a general-speech subspace from coarse-token prediction hidden states, then projects updates into its null space to minimize impact on general speech.The projection is illustrated as applying corrections in directions orthogonal to the general-speech subspace.Source: §3.3 · p. 5, Figure 2 · p. 5
3.4. Null-Space Constrained Knowledge Editing
- Given a target key and desired coarse-token value, a closed-form constrained update matches the target while approximately preserving K0 and is applied only to localized value projections.The procedure uses two forward passes to extract keys and values, a precomputed projection matrix, and O(d^2) operations for the rank-1 update.Source: §3.4 · p. 6
4. Experiments
The experiments test one-shot pronunciation correction without acoustic-decoder retraining and assess whether edits preserve general speech behavior. SonoEdit achieves 2.8% Target-WER and 3.15% Global-WER, with Global-WER statistically indistinguishable from the original model.
4. Experiments
- The experiments test whether one-shot null-space edits correct low-resource proper nouns, preserve prosody, semantics, and fluency, and remain parsimonious without decoder retraining.The acoustic decoder is not retrained in the experimental regime.Source: §4 · p. 6
4.1. Implementation Details
- Evaluation uses Orpheus-TTS and Sesame-TTS, with HardNoun-300 covering 300 proper nouns across six languages and 3,000 test utterances.The metrics assess target pronunciation, preservation-set WER, speaker similarity, and perceptual quality.Source: §4.1 · pp. 6–7
4.2. Layer Sensitivity Analysis
- Causal mediation, probe accuracy, and gradient norms all identify layers 15–21 as most phonetic, making them suitable for surgical editing.The analysis reports mean and standard deviation across 3,000 test utterances.Source: §4.2 · p. 7, Figure 3 · p. 7, Table 1 · p. 8
4.3. Null Space Constrained Editing
- SonoEdit achieves 2.8% Target-WER while maintaining 3.15% Global-WER, statistically indistinguishable from the original model.Table 2 compares editing methods using Target-WER for correction, Global-WER for general-behavior damage, and SIM for voice preservation.Source: Table 2 · p. 8, §4.3 · p. 8
- Pronunciation ratings reach 91% correctness and 88% OOD generalization, while preserved speech shows a 0.1% WER increase and 0.3Hz F0 RMSE drift.The reported mel-spectrogram distances decrease by 68–69%, and MOS remains stable.Source: §4.3 · p. 8, Table 3 · p. 9, Table 4 · p. 9
4.4. Ablation and Efficiency Analysis
- Removing the null-space constraint degrades Global-WER from 3.15% to 9.84% and SIM from 0.99 to 0.87; SonoEdit uses seconds and zero parameters.SonoEdit matches ROME’s one-shot efficiency, while FFT takes hours and LoRA takes minutes.Source: §4.4 · pp. 8–9, Table 5 · p. 9
5. Conclusion
SonoEdit combines causal tracing with null-space constrained editing to correct pronunciation without retraining. On HardNoun-300, it reports 91% correction with 3.15% Global-WER and 0.99 speaker similarity, while noting limits tied to corpus shifts, token granularity, edit scale, and error type.
5. Conclusion
- SonoEdit reports 91% correction with 3.15% Global-WER and 0.99 speaker similarity on HardNoun-300, outperforming unconstrained methods.The framework uses causal tracing to localize pronunciation and null-space constrained editing to restrict updates to general-speech null space.Source: §5 · p. 9
- Its null spaces come from a representative speech corpus, so deployment-distribution shifts may weaken orthogonality guarantees.Other stated limits include coarse semantic-token tracing, uncertain scalability to massive edit batches, and reduced suitability for global systematic errors such as accent bias.Source: §5 · p. 9