Bridge-SR: Schrödinger Bridge Speech SR

• Bridge-SR is a speech super-resolution model that directly maps low-resolution waveforms to 48 kHz high-resolution outputs using a data-to-data Schrödinger Bridge process.
• It employs a lightweight 1.7M-parameter architecture operating entirely in waveform space, preserving low-frequency structure while recovering missing high-frequency details.
• Ablation studies demonstrate that optimized noise schedules, data scaling, and auxiliary spectral losses yield superior LSD and SI-SNR metrics compared to traditional diffusion-based methods.

Bridge-SR is a Schrödinger Bridge–based speech super-resolution system that directly maps a low-sampling-rate waveform to its high-sampling-rate counterpart through a data-to-data stochastic process in the waveform domain. It is formulated as an any-to-48 kHz speech SR model, uses the observed low-resolution waveform as an informative endpoint of the bridge rather than sampling from pure Gaussian noise, and is designed to improve both synthesis quality and inference efficiency with a lightweight network backbone (Li et al., 14 Jan 2025).

1. Definition and task setting

Bridge-SR addresses speech super-resolution in the waveform domain. The input is a low-resolution waveform $x_{\text{LR}}$, and the output is a high-resolution waveform $x_{\text{HR}}$ at 48 kHz. During training, the input sampling rate is uniformly sampled from 6 kHz up to 48 kHz, while evaluation focuses on $8$k, $12$k, $16$k, and $24$k $\to 48$k settings. The training LR signal is obtained from the HR waveform through a chain of signal-processing low-pass and down-sampling filters with randomized settings, so the model learns to reconstruct the full-band waveform from its band-limited version (Li et al., 14 Jan 2025).

The system is motivated by two limitations identified in earlier SR approaches. First, conditional diffusion SR typically follows a noise-to-data trajectory: the reverse process starts from Gaussian noise even though the low-resolution waveform already contains strong structural information about the target. Second, feature-space or vocoder-based approaches often require additional compression networks or decoders and can rely on substantially larger backbones. Bridge-SR instead uses a waveform-domain Schrödinger bridge so that the generation path connects low-resolution and high-resolution speech directly, preserving low-frequency structure and focusing model capacity on the missing high-frequency components (Li et al., 14 Jan 2025).

A common misconception is that Bridge-SR is simply a conditional diffusion model with alternative conditioning. Its defining difference is the trajectory itself: the method learns a data-to-data bridge rather than a noise-to-data denoising chain. In the formulation adopted here, the low-resolution observation is not an auxiliary side input attached to a standard diffusion prior; it is one endpoint of the generative process.

2. Schrödinger-bridge formulation

Bridge-SR uses a tractable Schrödinger Bridge model in continuous time. The reference SDE is written as

$dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$

and the bridge problem is posed as

$$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$

In the detailed construction, the endpoint laws are specified as

$$p_0=\mathcal{N}(x_{\text{HR}},\epsilon_0^2 I), \qquad p_T=\mathcal{N}(x_{\text{LR}},\epsilon_T^2 I),$$

and inference proceeds with the reverse-time dynamics from LR at $x_{\text{HR}}$0 back to HR at $x_{\text{HR}}$1 (Li et al., 14 Jan 2025).

Under the linear-Gaussian assumptions, the bridge has a closed-form marginal distribution

$x_{\text{HR}}$2

with $x_{\text{HR}}$3 and $x_{\text{HR}}$4. Training samples $x_{\text{HR}}$5 from this bridge and learns a network $x_{\text{HR}}$6 by minimizing

$x_{\text{HR}}$7

This is a denoising regression objective over the bridge path, conditioned on the LR waveform (Li et al., 14 Jan 2025).

Within the broader theory of diffusion bridges, Bridge-SR is naturally situated on the Schrödinger-bridge side rather than the deterministic flow-matching side. A unified treatment of bridge problems describes super-resolution as a conditional bridge between an initial distribution $x_{\text{HR}}$8 and a target distribution $x_{\text{HR}}$9, implemented either as an ODE bridge or an SDE bridge. In that taxonomy, a Schrödinger-bridge SR system uses stochastic dynamics

$8$0

and exploits Brownian-bridge structure and path-space regularization (Kim, 27 Mar 2025).

3. Architecture and sampling procedure

Bridge-SR uses a lightweight $8$1M-parameter architecture following NU-Wave2 and WaveNet-style designs. The network takes as input the current bridge state $8$2, the time index $8$3, and the low-resolution waveform $8$4, and outputs an estimate of the high-resolution waveform $8$5. The model operates entirely in waveform space; no STFT or MDCT representation is used in the generative backbone (Li et al., 14 Jan 2025).

Sampling begins from the low-resolution waveform aligned to the 48 kHz temporal grid and scaled as required by training. The reverse process then integrates from $8$6 to $8$7 with either reverse-time SDE solvers or PF-ODE solvers adopted from Bridge-TTS. For 50-step and 8-step sampling, the paper uses a linear time schedule between $8$8 and $8$9 with a first-order PF-ODE sampler. Few-step settings are also reported: a 4-step second-order SDE sampler with $12$0, a 2-step first-order PF-ODE sampler with $12$1, and a 1-step PF-ODE sampler with $12$2 (Li et al., 14 Jan 2025).

The architectural point is not novelty in backbone design but compatibility between a small waveform predictor and a bridge process whose intermediate states remain mixtures of LR and HR content plus controlled noise. This suggests that the efficiency gains of Bridge-SR stem primarily from the generative path design rather than from scale alone.

4. Noise schedule, data scaling, and auxiliary objectives

Bridge-SR emphasizes three design choices: the bridge noise schedule, data scaling, and auxiliary spectral losses. The diffusion coefficient is parameterized as

$12$3

Among the tested schedules, the paper distinguishes Bridge-$12$4, Bridge-$12$5, and Bridge-SVP. Bridge-$12$6 is asymmetric, with $12$7 and larger $12$8, and allocates more noise to later timesteps while preserving low-frequency consistency because the implementation uses $12$9. In ablations for $16$0k$16$1k SR, Bridge-$16$2 yields LSD $16$3, Bridge-$16$4 yields $16$5, and SVP yields $16$6 (Li et al., 14 Jan 2025).

Data scaling is introduced because high-frequency waveform components have much smaller amplitude than low-frequency components. The model uses

$16$7

and applies it to both HR and LR signals; in the experiments, $16$8. This scaling is reported as crucial. Without scaling and without auxiliary loss, LSD degrades to $16$9 and SI-SNR drops to $24$0, whereas the full Bridge-$24$1 configuration reaches LSD $24$2 and SI-SNR $24$3 (Li et al., 14 Jan 2025).

The model is first trained for $24$4M steps with the bridge loss alone, then fine-tuned for $24$5k additional steps with auxiliary multi-scale STFT magnitude and anti-wrapping phase losses: $24$6 where $24$7 and $24$8. Removing these auxiliary losses degrades LSD from $24$9 to $\to 48$0, LSD-HF from $\to 48$1 to $\to 48$2, and SSIM from $\to 48$3 to $\to 48$4 (Li et al., 14 Jan 2025).

These ablations clarify that Bridge-SR is not only a bridge model in name. Its performance depends on preserving low-frequency consistency along the trajectory, rescaling the training target so that high-frequency reconstruction is not numerically suppressed, and supplementing waveform regression with spectrally structured objectives.

5. Empirical performance and efficiency

Bridge-SR is evaluated on the VCTK corpus, with approximately $\to 48$5 sentences from $\to 48$6 speakers resampled to 48 kHz. Training uses a window length of $\to 48$7 samples, batch size $\to 48$8, learning rate $\to 48$9, noise schedule $dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$0, $dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$1, and the $dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$2M-parameter backbone (Li et al., 14 Jan 2025).

For $dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$3k$dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$4k SR, Bridge-SR attains the best full-band LSD among the listed learned models and the best LSD-HF, while matching NU-Wave2 in SI-SNR.

Setting	Bridge-SR	Comparison
$dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$5k$dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$6k LSD	0.848	NU-Wave2: 0.927
$dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$7k$dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$8k LSD-HF	1.028	best reported
$dx_t = f(x_t,t)\,dt + g(t)\,dw_t, \quad t\in[0,T],$9k$$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$0k SI-SNR	25.04	NU-Wave2: 25.03
$$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$1k$$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$2k LSD	0.716	NU-Wave2: 0.740
$$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$3k$$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$4k LSD	0.928	NU-Wave2: 1.015
$$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$5k$$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$6k LSD	1.015	NU-Wave2: 1.140

The efficiency claim is especially notable in the few-step regime. For $$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$7k$$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$8k, NU-Wave2 with 8 steps achieves LSD $$\min_{p \in \mathcal{P}_{[0,T]}} D_{\mathrm{KL}}(p \parallel p^{\mathrm{ref}}) \quad \text{s.t.} \quad p_0 = p_{\mathrm{prior}},\; p_T = p_{\mathrm{data}}.$$9. Bridge-SR with 8 steps reaches $$p_0=\mathcal{N}(x_{\text{HR}},\epsilon_0^2 I), \qquad p_T=\mathcal{N}(x_{\text{LR}},\epsilon_T^2 I),$$0, and Bridge-SR with 4 steps reaches $$p_0=\mathcal{N}(x_{\text{HR}},\epsilon_0^2 I), \qquad p_T=\mathcal{N}(x_{\text{LR}},\epsilon_T^2 I),$$1. The paper therefore highlights that 4-step Bridge-SR already surpasses the 8-step conditional diffusion counterpart, while 50-step Bridge-SR gives the best overall quality at LSD $$p_0=\mathcal{N}(x_{\text{HR}},\epsilon_0^2 I), \qquad p_T=\mathcal{N}(x_{\text{LR}},\epsilon_T^2 I),$$2 (Li et al., 14 Jan 2025).

The sampling-step ablation also shows a characteristic difference from the diffusion baseline: Bridge-SR quality continues to improve as the number of reverse steps increases, with the ordering $$p_0=\mathcal{N}(x_{\text{HR}},\epsilon_0^2 I), \qquad p_T=\mathcal{N}(x_{\text{LR}},\epsilon_T^2 I),$$3, whereas diffusion quality is described as saturating earlier. This supports the interpretation that the bridge path remains informative throughout the reverse trajectory.

6. Position in bridge-based super-resolution research

Bridge-SR established a waveform-domain Schrödinger-bridge formulation for speech SR with explicit gains in quality-per-step and model compactness. Its central idea is that SR should be treated as HR$$p_0=\mathcal{N}(x_{\text{HR}},\epsilon_0^2 I), \qquad p_T=\mathcal{N}(x_{\text{LR}},\epsilon_T^2 I),$$4LR bridging rather than as conditional denoising from an uninformative Gaussian initial state. In that sense, it exemplifies a more general bridge-based generative modeling paradigm in which endpoint structure is task-specific rather than fixed to data versus noise (Kim, 27 Mar 2025).

Subsequent work extends the same bridge intuition into latent space. “Audio Super-Resolution with Latent Bridge Models” introduces latent bridge models for audio SR, where the waveform is compressed into a continuous latent space and the bridge becomes latent-to-latent rather than waveform-to-waveform. That work explicitly positions latent bridge models as a way to exploit the instructive prior information contained in the LR waveform, extends bridge-based SR beyond speech to audio and music, and reports any-to-48 kHz and any-to-192 kHz results through frequency-aware and cascaded latent bridges (Li et al., 22 Sep 2025). This suggests a trajectory within the literature from Bridge-SR’s direct waveform bridge to broader latent bridge frameworks.

Bridge-SR’s limitations are also clear from the reported scope. The experiments are conducted on clean single-speaker speech in VCTK; robustness to noisy or reverberant real-world audio, non-speech signals, and more extreme degradations is not fully addressed. The method still requires iterative SDE or ODE integration, and the Schrödinger-bridge framework demands careful schedule and scaling design. Even so, within its stated domain, Bridge-SR marks a precise shift in SR methodology: it treats low-resolution speech as a generative endpoint, preserves low-frequency structure throughout the path, and demonstrates that few-step bridge sampling can outperform an 8-step conditional diffusion baseline using the same order of backbone size (Li et al., 14 Jan 2025).