Flow-matching generators turn random noise into audio by following a learned velocity field along nearly straight paths for one to a few dozen steps. Systems such as VoicePrompter need only 1 to 6 steps, and SAM Audio needs 2 to 16. Nobody can currently say how often such a model will produce a bad output: a clipped frame, a voice that drifts away from the requested speaker, or a burst of energy. The verification community has tools for exactly this question, called reachability analysis: methods that bound every state a continuous system can reach from a set of starting points. The best statistical tools of this kind, Stochastic Lagrangian Reachability and GoTube, cost exponentially more as the number of dimensions grows. They stop at about 16 dimensions, while audio latents have hundreds. This is the continuous version of the state-space explosion that model checkers face with concurrent programs.
We propose to stop trying to cover the space of noise inputs. A generator's input is not an adversarial ball; it is a standard Gaussian draw. For Gaussian inputs, any function that changes slowly with its input concentrates around its typical value, at a rate that does not depend on dimension. If we can estimate how sensitive the end-to-end sampler is (noise in, safety score out), we get a tail bound on failure probability that needs no grid, no covering, and no rare-event simulation. Near-straight flow-matching paths are the regime where that sensitivity should be small and measurable. The two-sentence version: verify the generator against the noise distribution it actually samples from, not against a worst-case region. Gaussian concentration then turns an exponential covering problem into a sensitivity estimate whose cost does not grow with dimension.
Core Claim
Let $\Phi_c$ be the full sampler for condition $c$, meaning the ODE solver run from noise $x_0 \in \mathbb{R}^D$ to a generated latent. Let $f$ be a scalar safety score computed from the output, such as cosine similarity between a speaker embedding of the generated audio and the prompt speaker. Let $A$ be a set of Gaussian mass at least $1-\delta$, such as the thin shell where Gaussian samples live. If $f\circ\Phi_c$ has Lipschitz constant $L$ on $A$ (the largest rate at which the score can change per unit of input), we can extend it off $A$ without increasing $L$. The Gaussian concentration inequality then gives
$$ \Pr_{x_0\sim\mathcal{N}(0,I)}\big[f(\Phi_c(x_0)) < m - s\big] \;\le\; \delta + \exp\!\left(-\frac{s^2}{2L^2}\right), $$
where $m$ is the median score and $s$ is the margin to the failure threshold. The dimension $D$ does not appear. The only quantities to estimate are $m$, which Monte Carlo handles easily, and $L$ on the typical set.
We claim three things. First, for straight-path flow models, a trajectory-anchored estimate of $L$ is small enough to give non-vacuous certificates at failure rates of $10^{-4}$ to $10^{-6}$ with a few thousand sampler runs. Plain Monte Carlo would need millions of runs to reach the same rate. Second, the bound becomes looser as the paths become more curved. That makes straightness, which the field has optimized only for speed, also a measure of how verifiable a model is. Third, the certificate is statistical in exactly one place, the estimate of $L$, and we can control that step with the same kind of statistical Lipschitz caps GoTube uses.
Relation to Prior Work
Two papers supply the method. On the Verification of Neural ODEs with Stochastic Guarantees introduced the Lagrangian view. It bounds reachable sets by measuring sensitivity back to the initial set at each time, rather than pushing overapproximations forward step by step, which inflates them (the "wrapping effect"). It also showed that forward-mode sensitivity is roughly 50% faster than reverse mode. GoTube turned this into statistically estimated local Lipschitz caps around GPU-sampled trajectories and reached 16-dimensional models. We reuse the Lagrangian sensitivity computation and the statistical cap estimation.
These papers also motivate and constrain the design. Both certify worst-case enclosures of a ball of initial states, and the file reports that the cost of doing so is exponential in $D$. That design choice cannot scale to generative latents, so we change the question from "all inputs in a ball" to "Gaussian-probable inputs." The deterministic hybrid-system tools reviewed in Running SpaceEx on the ARCH14 Benchmarks are a further warning. Their formal analysis of the motor-transmission system failed after six collisions because error accumulated, and the fallback was trajectory sampling. That supports our choice to anchor every estimate to sampled trajectories.
On the generation side, the optimal-transport straight-line path used in work such as Fake it to make it, P-Flow, and E2 TTS is the property we expect to make $L$ small. A perfectly straight flow is close to a fixed linear map from noise to output, and a linear map's Lipschitz constant is just its largest stretch factor. The failure modes we target come from the file. CoVoMix2 names identity drift and disfluency-induced misalignment as open problems for non-autoregressive flows. AnyAccomp shows flow models can collapse completely off-distribution, with accompaniment adherence of 0.000 for its baseline on clean vocals. These papers measure average quality. None of them bounds tail failure rates, and that is the gap we address.
The concurrency side of state-space reduction, including value-centric partial order reduction, unfoldings, and quotienting, plays a narrow role here. It informs how we treat the discrete condition set, as described in the research program, and we do not claim more than that. CoVoMix2's parallel per-speaker tracks already remove the interleaving explosion in dialogue generation, so dressing that up in Petri-net language would add nothing.
Research Program
We will train an optimal-transport flow-matching text-to-speech model on LibriTTS, using mel or codec latents with $D$ ranging from 64 to 1024 by varying the latent bottleneck. We will train sibling models that share the architecture but use a curved variance-preserving diffusion path, so path curvature can be varied while everything else is held fixed. Samplers will run with Euler integration at 1, 2, 4, 8, and 32 steps.
The safety scores will be functions that can be differentiated through the generator. The first is speaker similarity to the prompt, which measures identity drift. The second is maximum per-frame log-energy after the vocoder, which measures clipping and bursts. The third is a chroma-alignment score for an accompaniment variant trained in the style of AnyAccomp.
For each condition $c$, the certificate pipeline has four steps. It samples noise inputs, restricts them to the Gaussian shell $A$ with an analytically known $\delta$, and computes Jacobian-vector products of $f\circ\Phi_c$ along each trajectory in forward mode. It then fits an extreme-value tail to the observed local gradient norms, plus GoTube-style ball caps around each sample, to obtain an upper estimate of $L$ that holds with probability $1-\gamma$. Finally, it plugs $m$ and $L$ into the bound above.
Conditions are a discrete set: speakers times texts, which in dialogue becomes a combinatorial product. Rather than certifying every combination, we test whether $L$ is controlled by a small set of factors. If certificates computed per speaker and per text-length bucket upper-bound the joint certificates on held-out pairs, the condition space can be decomposed. That is the one place where the compositional ideas from verification are used, and we treat it as an empirical question rather than an assumption.
Our theoretical contribution is a bound on the end-to-end $L$ for a $K$-step Euler sampler in terms of how much the learned velocity field changes along the path. For an exactly straight flow, this bound reduces to the stretch of a single linear map. This explains why few-step straight models should certify well and gives a training-time regularizer we can test.
Baselines
The naive reference point is plain Monte Carlo. We run the sampler $n$ times, count failures, and report an exact one-sided binomial upper confidence bound on the failure rate. This bound is valid with no assumptions. With zero failures observed, it is roughly $3/n$ at 95% confidence, so certifying $10^{-5}$ needs about 300,000 full generations. We give Monte Carlo exactly the same number of sampler evaluations (forward passes plus Jacobian-vector products, counted in GPU-seconds) as our method. Our method beats it if it reports a strictly smaller valid upper bound at matched compute. Validity is checked against ground-truth failure rates measured with ten million samples on the 64- and 128-dimensional models.
The strongest existing verification method is GoTube, run as published: a reachtube over a ball of initial noises around the typical set, with confidence $1-\gamma$ and tightness $\mu$, using the same JAX sampling budget. For comparison, the file reports that it reached 16-dimensional continuous-time RNNs and kept tight tubes for over 40 seconds on Dubins Car, while deterministic tools blew up before 20 seconds. No reported result was found on Lacuna for GoTube, Stochastic Lagrangian Reachability, or any reachability method on generative flow-matching or audio models. We therefore expect to establish this baseline ourselves. We will report the dimension at which it times out or its tube covers the failure threshold. Our method beats it if it gives a non-vacuous certificate at dimensions where GoTube does not.
There are two ablations. The first replaces our trajectory-local estimate of $L$ with the deterministic global bound, the product of layer spectral norms over all solver steps. This isolates whether local sensitivity estimation is what makes the bound useful; we expect the global version to be vacuous. The second swaps the straight-path model for its curved sibling at matched quality. This isolates whether straightness is the property that makes certification possible. On the generation side, CoVoMix2's reported DailyDialog speaker similarity of 0.56 against 0.40 for MoonCast gives the scale for choosing a meaningful identity-drift threshold in the dialogue extension.
Why not just…?
Why not just run more Monte Carlo samples? For moderate failure rates this is the right answer, and we will say so. Our method only helps in the tail. Certifying failure rates at $10^{-5}$ or below for every speaker in a deployed voice-cloning system multiplies hundreds of thousands of generations by thousands of conditions. The concentration bound turns a few thousand samples into a tail estimate. The real threat is that $L$ is large enough to make our bound looser than Monte Carlo at feasible budgets. The matched-compute comparison is designed to show directly where the crossover lies. If no crossover appears at $10^{-4}$ on the straight 8-step models, the claim fails.
Why not just fit an extreme-value distribution to the observed safety scores? Extrapolating the tail of $f$ directly is cheaper and does not depend on dimension either. However, it assumes a parametric tail shape for the output, with no structural reason for that shape to hold. Our bound puts the extrapolation into $L$, which is the gradient size of a known, smooth composite map and can be checked with ball caps around samples. We include direct extreme-value fitting on scores as an extra comparison at matched compute. If it is equally valid and tighter on held-out ten-million-sample ground truth, the structural argument does not pay for itself, and we will report that.
Why not just regularize training for straightness or robustness and skip certification? Straightness regularization is cheap and probably reduces failures. It produces no number a deployer can rely on, though. Without a tail estimate, nobody can tell whether a 1-step model fails ten times more often than an 8-step one. The curvature ablation measures how much straightness alone moves the true failure rate. Separately, the certificate measures whether the improvement shows up in a computable bound.
Proposed Experiments and Decisive Checks
The experiments run on LibriTTS single-speaker and zero-shot voice cloning. They extend to two-speaker dialogue using DailyDialog scripts, following CoVoMix2's evaluation setup with SA-SIM and SA-WER as quality guards. For the accompaniment model, they extend to MUSDB18 and MoisesDB clean-vocal conditions, where AnyAccomp reports adherence of 0.710 and 0.203 respectively. We report certified upper bounds against true failure rates, the empirical rate at which the bound is violated (which must stay at or below $\gamma$), compute cost, and how all of these scale with $D$ and step count.
Three outcomes would decide the question. The first is validity: across all models and conditions, the certified bound must exceed the ten-million-sample ground truth in at least a $1-\gamma$ fraction of runs. The second is dimension independence: at fixed step count, the certificate's tightness should stay roughly flat from $D=64$ to $D=1024$, while GoTube's cost grows exponentially. The third is the straightness link: the certified bound should loosen as measured path curvature increases, and the same ordering should hold for the true failure rate.
The proposal is falsified in two cases. If the estimated $L$ on the typical shell grows with $D$ for straight flows, the dimension-free claim fails. If validity breaks because local gradient norms miss rare high-sensitivity regions, the statistical step fails, and an honest negative result would show the limits of trajectory-anchored verification for generative models.
References
- VoicePrompter: Robust Zero-Shot Voice Conversion with Voice Prompt and Conditional Flow Matching
- SAM Audio: Segment Anything in Audio
- On the Verification of Neural ODEs with Stochastic Guarantees
- GoTube: Scalable Statistical Verification of Continuous-Depth Models
- Running SpaceEx on the ARCH14 Benchmarks
- Fake it to make it: Using synthetic data to remedy the data shortage in joint multi-modal speech-and-gesture synthesis
- P-Flow: A Fast and Data-Efficient Zero-Shot TTS through Speech Prompting
- E2 TTS: Embarrassingly Easy Fully Non-Autoregressive Zero-Shot TTS
- CoVoMix2: Advancing Zero-Shot Dialogue Generation with Fully Non-Autoregressive Flow Matching
- AnyAccomp: Generalizable Accompaniment Generation via Quantized Melodic Bottleneck
- Value-centric dynamic partial order reduction
- Using Unfoldings to Avoid the State Explosion Problem in the Verification of Asynchronous Circuits
- Compositional Model Checking of Real Time Systems
- LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech