Gepard 1.1

This release is a GRPO pass over gepard-1.0. It is a robustness release, not a quality release. The model drops and invents fewer words, and the rate of unusable output falls by roughly an order of magnitude. Voice cloning is unchanged and perceived audio quality is flat to marginally better — both measured, both reported below in What changed in 1.1.

GEnerative, Prosody-aware, Autoregressive text-to-speech model for Realtime Dialogue

Gepard is a text-to-speech model built for real-time conversation. It starts speaking the moment text begins arriving, generating audio piece by piece instead of waiting for a full sentence — so it feels like a live voice, not a recording. It's a single language model that learned text and speech together, so the output carries natural rhythm and timing rather than the flat, stitched tone of older pipelines.

The name evokes "Gepard"(/geh-PART/), German for cheetah — a nod to the model's low-latency, high-throughput streaming.

Want to use Gepard in production without hosting it yourself? You can skip the deployment and use our real-time TTS API — a fully managed, Cartesia-compatible service with millisecond time-to-first-chunk, voice cloning, and streaming built in. If you've used Cartesia, you already know how to use it: point the base URL at https://api.nineninesix.ai and the official Cartesia SDKs just work. Roughly 22 hours of audio for $5, and free credits to start (no credit card required). Try the live demo or grab an API key.

Highlights:

One clean pass per frame — the whole audio frame (32 orthogonal FSQ channels) is sampled in one step, no depth-transformer.
Extra quality, for free: CFG refinement (normally a two-pass cost) is baked into the weights, with the two-pass mode still available as a quality dial.
Real-time on vLLM: ~25× real time on a single RTX 5090, with first audio chunk (TTFA) in ~50 ms
Scales to many callers: One 96GB GPU (RTX Pro 6000 Blackwell) holds up to 256 conversations in parallel.
Voice cloning from a short clip: A few seconds of reference audio captures the speaker once, up front — cloning adds nothing to the per-word cost.

Languages

English: US, UK
Spanish: Mexico
Portuguese: Brazil
Dutch: Netherlands

Samples

Try the model live in the Gepard Demo Space.

Language / Accent	Sample
English (en-US)
English (en-GB)
Spanish (es-MX)
Portuguese (pt-BR)
Dutch (nl)

Model Details

Developed by: nineninesix.ai
Model type: Autoregressive (decoder-only) text-to-speech model
Backbone: Qwen3.5 full-attention transformer (14 layers, hidden 1024, 8 heads; ~500M params)
Audio codec: NVIDIA NeMo NanoCodec — FSQ, 22.05 kHz, 21.5 frames/s, 1.89 kbps
Parameters: ≈ 555.7M (backbone + audio interface + voice-cloning compressor)
Sample rate: 22,050 Hz
Languages: English, Spanish (es-MX), Portuguese (pt-BR), Dutch — plus five English accents
Training stage: GRPO on top of gepard-1.0
License: Apache 2.0 (the codec has NVIDIA Open Model License Agreement)

Benchmarks

Measured on the public Seed-TTS-eval — 1088 paired prompts, identical UUIDs and texts across every model.

Model	WER ↓	SIM ↑	UTMOS ↑	NISQA-MOS ↑	NOI ↑	COL ↑	DIS ↑
VoxCPM2	0.015	0.867	2.42	3.97	3.86	3.96	4.30
Fish-S2	0.016	0.789	2.80	4.18	3.87	4.14	4.44
OmniVoice	0.016	0.848	2.63	4.17	4.14	4.13	4.44
Qwen3-TTS	0.017	0.833	2.87	4.18	3.89	4.14	4.43
Echo-TTS	0.022	0.824	2.60	4.08	3.78	4.07	4.36
Gepard 1.0 (Ours)	0.036	0.585	2.64	4.25	4.16	4.16	4.51
Chatterbox	0.063	0.796	2.70	4.19	4.12	4.12	4.46

Gepard leads on perceived audio quality — highest naturalness (NISQA-MOS) and cleanest on noise, coloration, and discontinuity. In exchange for its streaming-first design and speed, it trades some speaker similarity (SIM) and word accuracy (WER) — a strong fit where a clean, natural real-time voice matters more than exact voice matching.

This table is the 1.0 row: 1.1 was not re-run on Seed-TTS-eval. What 1.1 changed relative to 1.0 was measured on a different benchmark and is reported directly below, so the two tables are not comparable to each other.

What changed in 1.1

Both checkpoints were run against the Multilingual Speech Benchmark v2.0 on 2026-09-02: 4 languages × 42 reference voices, 63,000 target utterances each, identical texts and identical prompt speakers. The two runs share a ruler — same recogniser and weights, same decoding, same benchmark version, same bootstrap seed and cluster column — so every delta below is a paired cluster bootstrap over speakers (2,000 replicates, 95 % CI), not two independent numbers placed side by side.

1. Intelligibility and robustness — where the release earns its number

measure	1.0	1.1	delta	95 % CI
WER, paired corpus	0.0801	0.0772	−0.0030 (−3.7 %)	[−0.0046, −0.0014]
WER, end-to-end ¹	0.0813	0.0773	−0.0040	—
unusable outputs (of 63,000)	88	12	−76	—
deletion rate — words dropped	0.01195	0.01111	−7.1 % rel.	[−0.00127, −0.00043]
insertion rate — words invented	0.01311	0.01186	−9.5 % rel.	[−0.00260, +0.00008]
substitution rate — wrong word	0.05505	0.05419	−1.6 % rel.	[−0.00147, −0.00030]
exact-match utterances	61.75 %	62.50 %	+0.76 pp	[+0.45, +1.06]
catastrophic, WER > 50 %	3.13 %	2.97 %	−0.16 pp	[−0.31, −0.01]
runaway clips, > 2× duration	200	188	−12	—
truncated clips, < 0.5× duration	595	557	−38	—
longest clip produced	29.8 s	22.9 s	−6.9 s	—

¹ every requested utterance charged to the model that produced it, failures counted as full deletions. Each model's published WER silently excludes its own worst output, so this is the row to quote for a production decision.

Read the error composition, not the headline. Substitutions barely moved — pronunciation is the same. Deletions and insertions fell by 7 % and 9.5 % relative: those two are the autoregressive failure signature, truncation and looping. GRPO suppressed the failure modes; it did not sharpen pronunciation.

2. Per language

language	n	WER 1.0	WER 1.1	delta	95 % CI	human anchor	1.0 vs anchor	1.1 vs anchor
all	62,902	0.0801	0.0772	−0.0030	[−0.0046, −0.0014]	0.0683	1.17×	1.13×
en-US	35,961	0.0700	0.0675	−0.0026	[−0.0046, −0.0009]	0.0774	0.91×	0.87×
es-MX	7,492	0.0802	0.0760	−0.0042	[−0.0110, +0.0011]	0.0613	1.31×	1.24×
nl-NL	10,497	0.0637	0.0626	−0.0011	[−0.0034, +0.0011]	0.0385	1.66×	1.63×
pt-BR	8,952	0.1575	0.1510	−0.0065	[−0.0141, +0.0010]	0.0738	2.13×	2.05×

human anchor is what the same recogniser scores on a real human recording of the same sentence; vs anchor is the model divided by it, so below 1.00× means fewer errors than the human recording. That ratio, not the raw WER, is the number to carry between benchmarks — the anchor varies more than 2× across these four languages, so 0.0626 on nl-NL is a worse result than 0.0675 on en-US.

Only the pooled figure and en-US clear the interval; es-MX and pt-BR show larger point estimates on 929 and 247 speakers, so their intervals are correspondingly wide. Across the 42 individual voices, 30 improved and 12 got worse — but zero regressions have an interval excluding zero, while four improvements do.

3. What did not change

axis	measure	1.0	1.1	delta	95 % CI
speaker similarity	SIM `wavlm_ft`	0.3914	0.3909	−0.0004	[−0.0011, +0.0002]
	SIM `ecapa`	0.4111	0.4105	−0.0006	[−0.0012, +0.0001]
	SIM `wavlm_sv`	0.8827	0.8824	−0.0003	[−0.0006, +0.0001]
naturalness	DNSMOS OVRL	3.087	3.088	+0.001	[−0.018, +0.022]
	UTMOS	2.477	2.526	+0.049	[−0.008, +0.100]
	NISQA MOS	4.411	4.464	+0.053	[−0.002, +0.115]
	NISQA distortion	4.394	4.477	+0.082	[+0.030, +0.139]

Voice cloning is unchanged. The SIM rows above are measured on the 62,912 clips both models produced successfully. Aggregate SIM over the full pool appears to rise, but that movement is carried entirely by the 88 clips 1.0 produced unusably — a speaker encoder embeds broken audio without complaint, and those rows sat in 1.0's mean at SIM ≈ −0.05. On the other 99.86 % of the pool, similarity is flat.

Naturalness is flat to marginally better. Distortion (nisqa_dis) is the one head that clears its interval, which is consistent with fewer degenerate segments rather than better synthesis. These are predicted MOS on a 420-clip sample clustering into 40 texts — directional only, and never to be read against the WER table.

Why the change is this narrow

GRPO modified 28 of the checkpoint's 296 tensors — the attention query and value projections in all 14 layers, and nothing else. The MLPs, the 32 codec heads, the embeddings, the stop head and the voice-cloning compressor are bit-for-bit identical to gepard-1.0. A small, attention-only adjustment is exactly the shape of change that suppresses truncation and looping without touching pronunciation or speaker identity, which is what the measurements show.

Limitations of this comparison

Generation wall-clock is not a controlled measurement. The two synthesis runs came from different directories with no guarantee of identical hardware, batching or concurrency. Treat any latency difference between 1.0 and 1.1 as unverified until a controlled A/B is run.
One recogniser. Whisper only. Its internal language model repairs some mispronunciation, and the repair is not guaranteed to be equal across the two models.
Naturalness has no human anchor — predicted MOS only, on 40 distinct texts.
Two of the benchmark's six subsets (ky, es-ES) were not evaluated for either model.
The benchmark contains no digits, no abbreviations and no long-form text (3–16 words), so text normalisation and long-context prosody are untested for both models.
GRPO training-data overlap with the benchmark was not audited. If any benchmark text or prompt speaker entered the GRPO stage, part of the improvement is contamination.

How these numbers compare to Seed-TTS-eval

Raw WER does not travel between benchmarks, and the two tables on this page are a good example. Both this benchmark and Seed-TTS-eval draw English from Common Voice and score with Whisper-large-v3 — yet the human anchor is 0.0774 here and 0.0214 there, 3.6× apart on the same corpus with the same recogniser. The difference is which rows were selected, not the ruler: this benchmark selects for full phoneme and diphone coverage and does not filter on how cleanly a recogniser reads the reference, so 40 % of its rows are ones where Whisper errs on the human recording. Restrict it to rows whose anchor matches Seed-TTS-eval's and gepard-1.0 scores 0.0436 instead of 0.0700, against 0.036 on Seed-TTS-eval — most of the apparent gap is the benchmark being harder by design. Prompts here are also shorter (2.5–5 s against 3–20 s) and the median target is 10 words. Text normalisation is not a factor: applying Whisper's EnglishTextNormalizer on top moves both the model and the anchor by −2.4 %.

So read the vs anchor column, not the WER, when placing these numbers beside anyone else's.

→ Full measurement record, including per-voice and per-length breakdowns, the failure accounting, the speaker-similarity pool split and the complete limitations: BENCHMARK_REPORT.md. Every number is recomputable from the published per-utterance parquet artifacts without a GPU.

What it's good for

Real-time and batch speech synthesis — voice agents, dialogue systems, content voiceover — in the supported languages, with optional zero-shot voice cloning from a short reference clip.

Keep in mind:

Quality is strongest in English; other languages vary by voice and content.
Real-time numbers come from the vLLM path (gepard-vllm); the reference PyTorch runner is the source of truth for behavior but isn't tuned for throughput.
Two-pass CFG dips voice similarity slightly versus single-pass — single-pass is the production default.
1.1 improves robustness, not voice cloning. If you selected 1.0 for its similarity behaviour, 1.1 will behave the same.

Use it responsibly. Don't clone a voice or synthesize speech without the speaker's consent, or create misleading or harmful content. You're responsible for following applicable laws and the licenses below.

Acknowledgments

Built on Qwen3-0.8B-Base as the backbone and NVIDIA NeMo NanoCodec for audio. Training data provided by the LAION team through the Emilia and EmoNet-Voice datasets.

Citation

If you use this work in your research, please cite:

@misc{gepard_2026,
  author        = {Pavlov, Denis and Abdurazakov, Ulanbek and Bakashov, Nursultan},
  title         = {{GEPARD}: Generative, Prosody-aware, Autoregressive Text-to-Speech Model for Realtime Dialogue},
  year          = {2026},
  eprint        = {2609.04222},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
  doi           = {10.48550/arXiv.2609.04222},
  url           = {https://arxiv.org/abs/2609.04222}
}

References

@misc{qwen3.5,
    title  = {{Qwen3.5}: Towards Native Multimodal Agents},
    author = {{Qwen Team}},
    month  = {February},
    year   = {2026},
    url    = {https://qwen.ai/blog?id=qwen3.5}
}


@inproceedings{kwon2023vllm,
  title={Efficient Memory Management for Large Language Model Serving with PagedAttention},
  author={Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Sheng, Ying and Zheng, Lianmin and Yu, Cody Hao and Gonzalez, Joseph E and Zhang, Hao and Stoica, Ion},
  booktitle={Proceedings of the 29th Symposium on Operating Systems Principles (SOSP)},
  pages={611--626},
  year={2023},
  eprint={2309.06180},
  archivePrefix={arXiv}
}

@article{dao2023flashattention2,
  title={FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning},
  author={Dao, Tri},
  journal={arXiv preprint arXiv:2307.08691},
  year={2023}
}

@article{nvidia2025nanocodec,
  title={NanoCodec: Towards High-Quality Ultra Fast Speech LLM Inference},
  author={Casanova, Edresson and Neekhara, Paarth and Langman, Ryan and Hussain, Shehzeen and Ghosh, Subhankar and Yang, Xuesong and Juki{\'c}, Ante and Li, Jason and Ginsburg, Boris},
  journal={arXiv preprint arXiv:2508.05835},
  year={2025}
}

@article{mentzer2023fsq,
  title={Finite Scalar Quantization: VQ-VAE Made Simple},
  author={Mentzer, Fabian and Agustsson, Eirikur and Tschannen, Michael and Malireddy, Srikanth and Alshina, Elena},
  journal={arXiv preprint arXiv:2309.15505},
  year={2023}
}

@article{nvidia2024magpie,
  title={Improving Robustness of LLM-based Speech Synthesis by Learning Monotonic Alignment},
  author={Neekhara, Paarth and Hussain, Shehzeen and Ghosh, Subhankar and Li, Jason and Valle, Rafael and Badlani, Rohan and Ginsburg, Boris},
  journal={arXiv preprint arXiv:2406.17957},
  year={2024}
}

@article{ho2022cfg,
  title={Classifier-Free Diffusion Guidance},
  author={Ho, Jonathan and Salimans, Tim},
  journal={arXiv preprint arXiv:2207.12598},
  year={2022}
}

@article{rafailov2023dpo,
  title={Direct Preference Optimization: Your Language Model is Secretly a Reward Model},
  author={Rafailov, Rafael and Sharma, Archit and Mitchell, Eric and Ermon, Stefano and Manning, Christopher D and Finn, Chelsea},
  journal={arXiv preprint arXiv:2305.18290},
  year={2023}
}

@article{meng2024simpo,
  title={SimPO: Simple Preference Optimization with a Reference-Free Reward},
  author={Meng, Yu and Xia, Mengzhou and Chen, Danqi},
  journal={arXiv preprint arXiv:2405.14734},
  year={2024}
}

@article{li2023blip2,
  title={BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models},
  author={Li, Junnan and Li, Dongxu and Savarese, Silvio and Hoi, Steven},
  journal={arXiv preprint arXiv:2301.12597},
  year={2023}
}

@inproceedings{khosla2020supcon,
  title={Supervised Contrastive Learning},
  author={Khosla, Prannay and Teterwak, Piotr and Wang, Chen and Sarna, Aaron and Tian, Yonglong and Isola, Phillip and Maschinot, Aaron and Liu, Ce and Krishnan, Dilip},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  volume={33},
  pages={18661--18673},
  year={2020},
  eprint={2004.11362},
  archivePrefix={arXiv}
}

@article{voicestar2025,
  title={VoiceStar: Robust Zero-Shot Autoregressive TTS with Duration Control and Extrapolation},
  author={Peng, Puyuan and Li, Shang-Wen and Mohamed, Abdelrahman and Harwath, David},
  journal={arXiv preprint arXiv:2505.19462},
  year={2025}
}

@article{radford2022whisper,
  title={Robust Speech Recognition via Large-Scale Weak Supervision},
  author={Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
  journal={arXiv preprint arXiv:2212.04356},
  year={2022}
}

@inproceedings{emilialarge,
  author={He, Haorui and Shang, Zengqiang and Wang, Chaoren and Li, Xuyuan and Gu, Yicheng and Hua, Hua and Liu, Liwei and Yang, Chen and Li, Jiaqi and Shi, Peiyang and Wang, Yuancheng and Chen, Kai and Zhang, Pengyuan and Wu, Zhizheng},
  title={Emilia: A Large-Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation},
  booktitle={arXiv:2501.15907},
  year={2025}
}

@article{emonet_voice_2025,
  author={Schuhmann, Christoph and Kaczmarczyk, Robert and Rabby, Gollam and Friedrich, Felix and Kraus, Maurice and Nadi, Kourosh and Nguyen, Huu and Kersting, Kristian and Auer, Sören},
  title={EmoNet-Voice: A Fine-Grained, Expert-Verified Benchmark for Speech Emotion Detection},
  journal={arXiv preprint arXiv:2506.09827},
  year={2025}
}

@article{chen2021wavlm,
  title={WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing},
  author={Chen, Sanyuan and Wang, Chengyi and Chen, Zhengyang and Wu, Yu and Liu, Shujie and Chen, Zhuoyuan and Li, Jinyu and others},
  journal={arXiv preprint arXiv:2110.13900},
  year={2021}
}