SonoEdit: Surgical TTS Editing

• The paper introduces SonoEdit, a surgical model editing framework designed to precisely correct TTS mispronunciations while preserving overall speech characteristics.
• It employs acoustic causal tracing to pinpoint critical transformer layers and applies a null-space constrained rank-one update to fix pronunciation errors.
• Experimental results show SonoEdit achieves low target-WER and high speaker similarity, outperforming full fine-tuning and other editing methods.

SonoEdit is a model editing framework designed for surgical correction of mispronunciations in pretrained neural text-to-speech (TTS) models. It enables one-shot, closed-form updates to pronunciation mappings—particularly for low-resource or out-of-distribution proper nouns—while provably preserving all other aspects of speech generation, such as prosody, fluency, and speaker identity. The technique introduces null-space constrained knowledge editing, leveraging causal tracing to localize parameter updates to specific transformer layers and employing null-space projectors to guarantee orthogonality to the subspace governing general speech production (Singh et al., 23 Jan 2026).

1. Motivation and Problem Scope

Modern LLM-based TTS architectures achieve high naturalness but consistently mispronounce proper nouns—especially those that are rare, non-English, brand, or geographic names—due to underrepresentation in monolingual pretraining corpora. Existing remediation approaches include:

• Multilingual data collection and annotation: Scalable data curation is cost-prohibitive.
• Supervised fine-tuning (including PEFT or LoRA): Requires ground-truth pronunciation data and risks catastrophic forgetting of unrelated behaviors.
• Manual phoneme injection: Demands expert phonetic labor and lacks generalization to unseen word forms.

These approaches therefore force undesirable trade-offs among cost, coverage, and the risk of damaging unrelated acoustic behaviors. There is pressing need for a technique that corrects isolated mispronunciations with zero retraining, no added parameters, and no collateral impact elsewhere in the model (Singh et al., 23 Jan 2026).

2. Technical Foundations and Methodology

SonoEdit integrates two core methodological advances:

• Acoustic Causal Tracing (ACT): Layer-localization procedure to identify which transformer layers encode the relevant grapheme-to-phoneme mapping for a mispronounced surface form.
• Null-Space Pronunciation Editing: A closed-form, rank-one parameter update to the localized weight matrices, strictly orthogonal to the principal subspace responsible for generic speech behaviors.

The core workflow proceeds as follows:

1. ACT Layer Selection: Systematically inject noise into text-encoder activations, then restore each layer’s clean hidden state and compute the “acoustic causal impact” on the coarse acoustic output token for the mispronounced word. The layers with highest impact define the parameter subset for editing.
2. Speech Manifold Characterization: Aggregate hidden state vectors at relevant coarse-token prediction steps across a general speech corpus to form matrix $K_0$, representing the preserved-behavior subspace.
3. Null-Space Projector Computation: Compute the orthogonal projector $P = I - UU^T$, where $U$ spans the principal components of $K_0K_0^T$ via SVD.
4. Rank-One Closed-Form Update: For a target instance with current key $k_*$ and desired value $v_*$, the solution is:

$$\Delta W = \frac{v_* - W k_*}{k_*^T P k_*} (P k_*)^T$$

Applied to the localized weights, this steers $Wk_*$ to $v_*$ while exactly preserving predictions for all directions in the span of $K_0$.

All steps are $O(d^2)$ per layer, requiring two forward passes and one SVD.

3. Mathematical Formulation

Let $W \in \mathbb{R}^{d \times d}$ be the target weight matrix. The constrained problem is:

• Target constraint: $(W + \Delta W) k_* = v_*$ (correct the specific mispronunciation)
• Preservation constraint: $\Delta W K_0 = 0$ (do not perturb other speech behaviors)

This is formalized as minimizing

$$\min_{\Delta W} \| (W + \Delta W) k_* - v_* \|_2^2 \quad \text{subject to} \quad \Delta W K_0 = 0$$

The null-space method guarantees the update lies in $S_p^\perp$, the orthogonal complement of generic speech behaviors, yielding the unique closed-form solution given above. Equivalently, update direction can be expressed as a null-space projection of the target loss gradient.

4. Algorithmic Summary

The following table summarizes the key computational steps:

Step	Operation	Output
1. Causal Tracing	Inject noise, calculate layerwise causal impact	Edit target layer indices $L_\text{edit}$
2. Null-Space Compute	Aggregate $K_0$, compute SVD, form $P$	Null-space projector $P$
3. One-Shot Edit	Compute $k_*, v_*, \Delta W$ as above	Edited weights $W + \Delta W$

All updates are performed only at the identified mid-to-late transformer layers governing G2P translation (Singh et al., 23 Jan 2026).

5. Experimental Design and Results

Experiments use the HardNoun-300 benchmark: 300 rare proper nouns from six languages (English, Spanish, French, German, Japanese, Hindi), each embedded in 10 contextual utterances. Orpheus-TTS is the base synthesis backbone. Key metrics include:

• Target-WER: Word error rate on just the corrected nouns.
• Phoneme Error Rate (PER): Phonemic transcription accuracy.
• Global-WER: WER on a holdout preservation corpus.
• Speaker Similarity (SIM): Cosine similarity of WavLM embeddings.
• MOS: Perceptual speech quality (UTMOS scale).

Results:

Method	Target-WER (%)	Global-WER (%)	SIM	MOS
Original	86	N/A	N/A	N/A
Full Finetune	2.1	18.45	N/A	N/A
LoRA (r=16)	4.5	5.12	N/A	N/A
ROME (unconstr.)	8.2	12.3	N/A	N/A
SonoEdit	2.8	3.15	0.99	4.18

Ablation of the null-space constraint increases global-WER to 9.84% and reduces SIM to 0.87, demonstrating that constrained orthogonality is essential for surgical editing (Singh et al., 23 Jan 2026).

6. Comparative Analysis and Related Methodologies

SonoEdit’s methodology is distinct from:

• Standard fine-tuning/PEFT approaches: Require iterative retraining and cannot guarantee preservation of unrelated speech behaviors, often introducing significant global degradation.
• ROM-based model editing: Unconstrained rank-one edits yield larger collateral damage.
• Manual phoneme annotation: Lacks generalization and scalability.

The approach provides closed-form solutions with formal preservation guarantees, unique among editing methods for sequence-to-sequence or acoustic models.

A plausible implication is that similar null-space constrained updates might be extended to cross-modal settings, such as embedding-based video-sound retrieval systems exemplified by Soundify (Lin et al., 2021), or to multimodal LLM knowledge editing, by leveraging embedding manifold structure and causal tracing for parameter localization.

7. Limitations and Perspectives for Extension

SonoEdit’s limitations include reliance on fixed-corpus null-space estimation—which may erode under dataset shift or domain adaptation—and focus on coarse token levels, potentially missing fine-scale acoustic variation. Scalability to batch edits is an open question: repeated rank-one updates could progressively saturate the null-space, diminishing preservation guarantees.

Potential future directions include:

• Batch simultaneous editing: Utilizing block-diagonal or low-rank null-space projectors for multi-word edits.
• Truly end-to-end TTS editing: Extending causal tracing and editing beyond coarse tokens to more granular acoustic representations.
• Application to multimodal/planner layers: Adapting the framework to cases where desired knowledge edits span modalities or require deeper planner/scheduler updates.

SonoEdit establishes a parsimonious, mathematically-grounded foundation for surgical model editing in neural TTS, achieving a state-of-the-art balance between targeted correction and global behavior preservation (Singh et al., 23 Jan 2026).