LauraTSE: Target Speaker Extraction

• LauraTSE is a generative framework for target speaker extraction that employs auto-regressive decoders and neural audio codec representations.
• The system integrates a Conformer-based feature encoder and an encoder-only vocoder LM to reconstruct high-quality speech from mixed inputs.
• A two-stage discriminative–generative pipeline refines coarse estimates, balancing perceptual enhancement with semantic fidelity as validated on Libri2Mix metrics.

LauraTSE is a generative framework for target speaker extraction (TSE), employing an auto-regressive (AR) decoder-only LLM in conjunction with neural audio codec representations. It is designed for the separation and high-fidelity reconstruction of speech from a specific target speaker, given a short enrollment utterance of that speaker, from a mixed audio input. In its most advanced configuration, LauraTSE is integrated into a two-stage discriminative–generative pipeline (USEF-Laura-TSE), which leverages the control and robustness of a discriminative front-end with the perceptual quality enhancement capabilities of a generative back-end (Zeng et al., 9 Jan 2026, Tang et al., 10 Apr 2025).

1. System Architecture

The core LauraTSE system consists of three main components:

(a) Feature Encoder (Conformer):

A Conformer-based encoder processes both the enrollment utterance ($r$) and the mixed speech input ($m$), generating continuous embeddings $E_r$ and $E_m$ from log-mel spectrogram inputs:

$E_m = \mathcal{C}(M_m),\; E_r = \mathcal{C}(M_r).$

(b) AR Decoder-Only LLM (Decoder-only LM):

Conditioned on $[bos; E_r; sep; E_m; tse]$, this component autoregressively predicts the discrete tokens for the first $n$ residual vector quantization (RVQ) codec layers of the target speech:

$$P_\theta(\hat{D}_n | E_m, E_r) = \prod_{i=1}^T P_\theta(\hat{D}_n^{(i)} | \hat{D}_n^{(1:i-1)}, E_m, E_r).$$

(c) Encoder-Only Vocoder LLM (Vocoder LM):

This one-step Transformer takes $[E_r, E_m, \hat{D}_n]$ and outputs a fine-grained embedding $\hat{E}_s$ representing the sum of all RVQ layers, which is then passed through a frozen neural audio codec decoder to produce the final waveform.

Two-Stage Discriminative–Generative Extension:

In the USEF-Laura-TSE framework, a discriminative front-end (USEF-TFGridNet) first extracts a coarse target speech estimate $m$0 through a sequence of operations: STFT, 2-D convolutional encoding, cross multi-head attention, concatenation, TF-GridNet processing, and decoding (transposed-conv + iSTFT). The generative LauraTSE back-end ($m$1) then refines this intermediate representation:

$m$2

treating $m$3 as the conditional context in place of $m$4 to reconstruct the final high-fidelity waveform (Zeng et al., 9 Jan 2026).

2. Mathematical Formulations and Training Losses

Generative Module

• AR LM (Decoder-only):

Standard cross-entropy loss over predicted codec tokens at each time step and across $m$5 layers:

$m$6

• Encoder-Only Vocoder LM:

Supervised with ground-truth summed embeddings, using an $m$7 loss:

$m$8

Discriminative Front-End

• Initially trained with complex-spectral MSE; during joint training, optionally includes an SI-SDR loss:

$m$9

where $E_r$0 and $E_r$1.

Combined Objective

• For joint training with an unfrozen front-end:

$E_r$2

where $E_r$3 aggregates the AR and vocoder losses.

3. Inference Procedures and Collaborative Training

• Pipeline:

1. Extract log-mel spectrograms for $E_r$4 and $E_r$5.
2. Encode with the shared Conformer to obtain $E_r$6.
3. Run the AR LM for coarse token prediction.
4. Map each token to its embedding, sum across layers for $E_r$7.
5. Invoke the encoder-only LM to compute $E_r$8.
6. Decode via neural audio codec to recover the time-domain waveform.

• Two-Stage Mode:

At inference, the AR LM’s free-running output may be partially or completely substituted with the discriminative front-end’s own codec-encoded tokens at an injection ratio $E_r$9, trading between fully autoregressive generation $E_m$0 and total reliance on front-end pseudo-labels $E_m$1.

• Front-End Training Variation:
  • Freeze: Discriminative module weights fixed; generative back-end learns to refine the fixed output.
  • Unfreeze: Gradients flow from generative losses into the discrimination module, enhancing consistency at the cost of more complex balancing, mediated through $E_m$2.

4. Experimental Results and Evaluation Metrics

Evaluation has been conducted primarily on the Libri2Mix clean test set, using the following metrics:

• DNSMOS (SIG, BAK, OVRL): Perceptual quality and noise suppression.
• NISQA: Overall non-intrusive speech quality assessment.
• SpeechBERT Score: Semantic consistency in self-supervised embedding space.
• dWER: Differential word error rate via Whisper ASR (semantic intelligibility).
• Speaker Similarity: Cosine similarity in WavLM-SV and WeSpeaker spaces.

Comparative results:

• The purely generative LauraTSE achieves OVRL ≈ 3.34, NISQA ≈ 4.33, dWER ≈ 0.159, speaker sim ≈ 0.97 (Zeng et al., 9 Jan 2026), showing high perceptual quality but weaker intelligibility and fidelity compared to discriminative models.
• The discriminative baseline (USEF-TFGridNet-L) has dWER ≈ 0.075 and high speaker similarity (>0.98) but lower OVRL ≈ 3.27, NISQA ≈ 4.32.
• The two-stage USEF-Laura-TSE-L (front-end unfrozen, SI-SDR loss, multiset training) achieves OVRL ≈ 3.32, NISQA ≈ 4.45, dWER ≈ 0.117, speaker sim ≈ 0.98, representing a balance between semantic consistency and perceptual enhancement.

Model	OVRL	NISQA	dWER	Speaker Sim
USEF-TFGridNet-L (D)	3.27	4.32	0.075	>0.98
LauraTSE (AR+vocoder) (G)	3.34	4.33	0.159	0.97
USEF-Laura-TSE-L (D+G)	3.32	4.45	0.117	0.98

D: Discriminative, G: Generative, D+G: Discriminative–Generative two-stage.

5. Ablation Studies and Architectural Insights

• RVQ Layer Depth ($E_m$3):

Varying $E_m$4 from 1 to 3 in the AR LM results in minimal differences in SIG/OVR metrics (≤0.03), indicating that even minimal coarse content suffices for downstream performance (Tang et al., 10 Apr 2025).

• Input Modality:

Discrete I/O, where Stage I operates purely on tokens rather than continuous embeddings, underperforms continuous-feature input modes. Replacing the Conformer with a frozen ASR Conformer or WavLM embeddings degrades speaker-related scores.

• Data Efficiency:

LauraTSE trained on 460 h LibriSpeech matches or outperforms other generative models trained on 5000 h multi-task data (e.g., AnyEnhance) in speaker and semantic similarity metrics.

• Training Regime:

In joint training, fine-tuning the discriminative front-end (as opposed to freezing) and applying SI-SDR loss with a tuned $E_m$5 improves semantic fidelity without perceptual degradation.

6. Limitations, Extensions, and Outlook

While LauraTSE establishes strong efficiency in single-task settings and shows effective performance with coarse-to-fine generation, several limitations are reported:

• Absence of explicit large-scale data scaling studies; current findings suggest data efficiency, but further quantification is needed.
• The two-stage coarse-to-fine structure increases system complexity and latency, suggesting a potential benefit for unified end-to-end sequence modeling.
• The neural codec is trained on clean speech only; adapting or jointly training the codec in noisy or mixture conditions may yield further robustness.
• Extending the backbone to multimodal or multi-task setups, such as noise suppression or dereverberation, is highlighted as a future direction (Tang et al., 10 Apr 2025).

• "Discriminative-Generative Target Speaker Extraction with Decoder-Only LLMs" (Zeng et al., 9 Jan 2026).
• "LauraTSE: Target Speaker Extraction using Auto-Regressive Decoder-Only LLMs" (Tang et al., 10 Apr 2025).