Audio8-TTS-Realistic: 0.6B voice cloning model, 23 languages

What changed

Edge0 announced Audio8-TTS-Realistic, an open-weight text-to-speech model for zero-shot voice cloning. The company released the 0.6B-parameter model with weights and code under Apache 2.0, along with a complete repository containing the checkpoint, 44.1 kHz audio codec, tokenizer, processor and custom inference code.

The key upgrade from its predecessor, Audio8-TTS-Preview-0.6b, is expanded language support. Audio8-TTS-Realistic now covers 23 languages, adding Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish and Vietnamese to the earlier 11-language lineup. The model retains the same Transformers inference interface and zero-shot voice cloning capability, requiring only one reference audio clip and its transcript to clone a voice.

The model uses a dual-autoregressive architecture inspired by Fish Audio S2 Pro. The slow branch contains 24 layers at width 896 and predicts text tokens plus the first audio codebook at approximately 21.5 frames per second. A shallow 4-layer fast branch then predicts the remaining nine codebooks in parallel. The bundled codec operates at 44.1 kHz and handles both reference-audio encoding and waveform decoding, supporting up to 2,048 packed text and audio positions of context.

Why it matters

The release addresses a practical gap in the open-source TTS landscape. Developers building voice applications now have an alternative to commercial TTS services, which charge subscription fees. The 601-million-parameter footprint fits on modest hardware: Edge0 reports peak memory usage of 1.1 to 1.2 gigabytes on an Apple M2 CPU using INT4 ONNX format, and the model can run on CUDA GPUs at reasonable inference cost.

The multilingual expansion matters for teams serving non-English markets. Adding 12 languages in a single release lets developers deploy voice cloning across broader geographic regions without licensing multiple proprietary systems. Accessibility to teams outside wealthy markets improves, since the Apache 2.0 license permits commercial and internal use without per-seat fees.

On standard benchmarks, Audio8-TTS-Realistic performs ahead of comparable open models. The company reports a 4.86% average error rate on CV3 across nine languages, outperforming Qwen3-TTS-1.7B at 7.73% and VoxCPM2 at 5.35%. On the MiniMax benchmark covering 24 languages, Audio8 achieves 5.89% average error. Larger commercial and research models still lead: Higgs TTS 3 reports 4.41% on CV3 and 2.74% on MiniMax. But for the model’s size class, the accuracy-to-parameter ratio is competitive.

The training approach differs from baseline supervised fine-tuning. Edge0 applied a learned realism critic in the final stage that scores outputs against matched human recordings. The company trained this critic on rankings from a private blind listening study across six open-source TTS systems, creating a quality ladder that penalizes synthetic artifacts like rhythm compression and unnatural pauses. This method improved error rates on lower-resource languages: their error fell from 22.53% to 20.77%, and speaker similarity rose from 0.637 to 0.679, even though training used only Chinese and English prompts. This suggests the critic captures prosodic qualities that generalize across languages.

What to test

Before adopting this model in production, validate these claims:

Accuracy claims on your data. Edge0 provides benchmark numbers on CV3, MiniMax and Seed-TTS datasets, but notes they differ from other evaluation pipelines and text normalization procedures. Run the model on your own test set in your target languages to confirm the error rates apply to your use case.

Voice quality on lower-resource languages. The company acknowledges weak performance on certain languages. Thai reaches 48.57% word error rate on CV3, and Cantonese hits 31.35% character error rate on MiniMax. If you need Thai, Turkish or Vietnamese at high accuracy, test locally before committing engineering effort.

Inference latency and throughput. Edge0 notes that throughput and streaming were not benchmarked. Test end-to-end latency on your target hardware using the SGLang Omni adapter if you need real-time or batch speech generation. A free Playground runs on Hugging Face ZeroGPU for initial experiments, but that setup will not reflect production constraints.

Fine-tuning stability. The repository includes supervised fine-tuning scripts. If you plan to adapt the model to domain-specific voices or accents, verify that fine-tuning converges on your data and does not degrade performance on the general case.

Reference audio quality requirements. The model needs a clean reference clip and its exact transcript for voice cloning. Test how sensitive clone quality is to background noise, volume variations and transcript accuracy. Edge0 does not document minimum reference audio duration or quality thresholds.

Memory footprint on target devices. The 1.1 to 1.2 GiB peak memory figure applies to M2 CPU inference with INT4 quantization. Measure actual memory use on your GPU or CPU, and confirm the deployment path if you need sub-gigabyte footprints. The source notes that no Realistic ONNX export is publicly available yet.

Common questions

How to clone a voice for TTS?

According to Audio8-TTS-Realistic documentation, voice cloning requires a clean reference audio clip and the exact transcript of that clip. The model uses these as input to the zero-shot voice cloning process. Edge0 provides a command-line interface and code examples in the GitHub repository for users to implement this workflow.

What is the most realistic TTS voice?

The source does not define realism or crown a single most realistic system. However, Edge0 argues that standard metrics like word error rate and speaker similarity miss prosodic qualities such as rhythm, pauses and timing. The company trained a realism critic on human-ranked recordings to optimize for these factors beyond typical benchmarks, which it claims improves naturalness in speech output.

The conclusion

Audio8-TTS-Realistic delivers a genuine tradeoff: significantly smaller parameter count than 4B-class alternatives with competitive accuracy on mid-range benchmarks and broader multilingual reach. For teams that need voice cloning on-device or behind a firewall, the Apache 2.0 license and 0.6B footprint remove cost and deployment friction compared to commercial systems.

The realism training approach is novel in its use of a learned critic scoring against human recordings, but remains unproven at scale. Independent evaluation by practitioners on real applications will determine whether the claimed prosodic improvements hold up outside the authors’ test set.

Watch for: (1) community contributions extending language coverage or improving lower-resource language accuracy, (2) published comparisons from independent evaluators on realistic voice cloning tasks beyond benchmarks, and (3) adoption in production voice applications that report real-world performance.

The model is immediately available on Hugging Face with a free Playground for initial experiments. Production deployment requires PyTorch 2.5 or newer and Transformers 4.57 or newer.

Related reading

• ElevenLabs vs PlayHT: AI Voice Cloning
• AssemblyAI-TTS-1: $15 per 1M characters for voice agents
• Anthropic OSS Scanner free vulnerability detection for open source