Tencent AuK Reality Check: 1.5B Speech Model, 4-Step Flash and Strong Vendor Benchmarks—Independent Reproduction Is Still Missing

Tencent has released AuK, an MIT-licensed speech-generation and editing model family, together with source code and downloadable weights. The release is unusually broad: one instruction interface covers zero-shot and instruction-controlled text-to-speech, content editing, acoustic and paralinguistic editing, enhancement, and source separation. The technical report was submitted to arXiv on September 8, 2026 at 15:59:16 UTC, while the current GitHub README marks the open-source code-and-weights release on September 9.

What actually shipped

Tencent describes AuK as a 1.5B-parameter foundation model. That headline should not be mistaken for the entire serving footprint. The released inference instructions also download Qwen2.5-Omni-3B as the multimodal semantic encoder, while the VAE is loaded separately at runtime. In other words, “1.5B” describes the AuK generative model rather than every parameter and component needed by the full runtime stack.

There are two variants. AuK is the full-quality model. AuK-Flash is a distilled checkpoint that uses 4 fixed function evaluations and no classifier-free guidance, versus 32 function evaluations with CFG 2.0 for the full model. The paper reports a 4.5× wall-clock speedup for Flash under the same hardware, output duration, and batch size. That is a controlled producer measurement, not a general guarantee about real-time factor, time-to-first-audio, concurrency, or end-to-end latency on arbitrary hardware.

The training description is also substantial. The authors report about 3.03 billion instruction–audio instances and 1.95 million hours of effective audio supervision across five task families. The model uses natural-language instructions plus optional audio context, a VAE acoustic latent space, and a hybrid rectified-flow Transformer. Editing receives human-preference post-training, while speech generation receives reward-based reinforcement learning.

The headline benchmark results are strong—but producer-run

The paper reports a wide benchmark suite, but the entire table is authored by the AuK team. Until an independent group reproduces the exact checkpoints, prompts, evaluators, baseline revisions and inference settings, these numbers should be treated as producer-reported evidence, not independent leaderboard confirmation.

On Seed-TTS-Eval, which measures zero-shot voice cloning across English, Chinese and difficult Chinese text, AuK reports an average recognition error of 2.65% and speaker similarity of 0.795. AuK-Flash reports 2.85% and 0.790. In the same paper table, Qwen3-TTS is shown at 3.07% average recognition error and 0.745 similarity, while Seed-TTS reaches 0.778 similarity. The authors say they run the generation evaluations three times and report the mean, using benchmark-specific instruction templates and duration estimation without the optional prompt enhancer.

For instruction-controlled synthesis on InstructTTSEval, AuK reports 83.37% Chinese DSD accuracy and 81.60% English DSD; AuK-Flash reports 78.80% and 82.40%. This is a good example of why a single “best model” label is too coarse: the full model is stronger on the Chinese split in the reported table, while Flash ties the best displayed English DSD result.

On MMAE-Speech, AuK reports 48.23% IFR, 88.11% CR, and 12.44% EMR. AuK-Flash reports 46.62%, 86.41%, and 13.85% EMR, respectively. IFR measures success on instruction-following rubrics, CR measures preservation of attributes not requested for editing, and EMR requires satisfying all instruction-following and preservation rubrics simultaneously. Flash leading EMR while the full model leads IFR and CR shows that even within one benchmark the trade-off depends on which metric matters.

On SpeechEditBench, the paper reports AuK at 91.83% content-editing joint success, 71.33% prosody, and 37.07% acoustic editing. AuK-Flash reports 39.25% on the paralinguistic category, above the displayed Step-Audio-EditX result of 31.25%. These are category-specific scores and should not be collapsed into a single universal editing rank.

Signal-restoration results show another split. On CHiME-4, AuK-Flash reports 7.84% WER and UTMOS 3.91, while AuK reports 7.98% WER and UTMOS 3.72. On Libri2Mix two-speaker separation, AuK reports 9.12% WER and 0.96 speaker similarity, while Flash reports 10.07% WER, the same 0.96 similarity, and higher reported UTMOS of 4.03 versus 3.87. The authors interpret this as the full model preserving linguistic content more strongly while Flash often scores better on predicted perceptual quality.

These benchmark families measure different things and use different automatic evaluators. WER/CER measures recognition accuracy, SIM measures speaker identity preservation, DNSMOS/UTMOS estimate perceptual quality, and editing suites use their own rubric or category success metrics. They should not be numerically ranked against one another.

Efficiency claims need the exact harness

AuK-Flash's most practical claim is the 4.5× speedup, but the denominator matters. The comparison is against the team's own 32-NFE, CFG-enabled AuK teacher, with equal hardware, batch size and output duration. It does not establish that Flash is 4.5× faster than other TTS or editing systems, nor does it supply a universal requests-per-second figure.

The public README does make deployment reproducible enough to inspect: it provides CLI, Python, Gradio and ComfyUI paths and documents BF16 inference. But the release materials reviewed here do not provide a standardized independent throughput table across GPUs, quantizations, batch sizes and audio durations. There is also no first-party metered API price to compare with hosted speech services. The weights are downloadable, so real cost depends on hardware, utilization, energy, storage and operations.

The code and AuK weights are released under the MIT license. The runtime also depends on separately downloaded components, including Qwen2.5-Omni-3B, whose own licensing and deployment requirements must be checked separately rather than assuming the AuK MIT license governs every dependency.

SWE-bench and Terminal-Bench are not applicable here

AuK is a speech generation and editing system, not a software-engineering agent. SWE-bench Verified, SWE-bench Pro and Terminal-Bench are therefore not applicable and are intentionally left unfilled. Importing scores from Qwen, Hunyuan or any unrelated text model would create a false comparison.

What independent evidence is missing

The key gap is simple: the current technical report and repository provide enough detail to reproduce the producer experiments, but a fresh bounded search did not find a credible third-party reproduction of the exact AuK or AuK-Flash checkpoint on the same benchmarks. No reliable exact-match X or Reddit technical evaluation was found either, so there is not enough evidence to claim public consensus.

Useful next tests would include a pinned independent Seed-TTS-Eval run, blinded human listening tests alongside automatic MOS predictors, per-task latency and VRAM measurements on named GPUs, ablations for the prompt enhancer and semantic encoder, and independent editing benchmarks that verify both requested changes and preservation of untouched speech characteristics.

Practical verdict

AuK is notable because it attempts to collapse a fragmented speech stack into one instruction-driven model rather than because one headline score settles the field. The released 1.5B AuK model plus separate Qwen2.5-Omni-3B semantic encoder, the 4-step AuK-Flash path, broad editing surface and permissive AuK license make it technically interesting for self-hosted experimentation.

The benchmark evidence is promising but still mostly first-party evidence from the release paper. The safest conclusion today is that AuK deserves independent reproduction—not that its launch table has already established a definitive speech-model leaderboard.

Primary sources: AuK technical report, Tencent-Hunyuan AuK repository, AuK-Flash weights.