Edge0's 0.6B multilingual speech synthesis model, an improved version of Audio8-TTS-Preview: it can clone voice timbre in a zero-shot manner given a reference recording and its transcript, or synthesize directly without a reference; it supports 23 languages including Chinese, Cantonese, English, Japanese, and Korean (12 more than Preview), and outputs 44.1 kHz audio.
Rising score 118 · Likes 119 · Downloads 813
Reasons
- Adopts the DualAR structure from Fish Audio S2 Pro: the slow autoregressive model predicts one semantic token per frame, while the fast autoregressive model generates 10 codebooks; the main model has 601M parameters (excluding the 44.1 kHz codec included with the repository), and the repository contains checkpoints, codecs, tokenizers, and code.
- Officially reported Seed-TTS error rates are 1.25 for Chinese and English (Preview: 1.23) and speaker similarity is 0.704 (Preview: 0.682); on the MiniMax 24 language set, the error rate is 5.89, which is better than Qwen3-TTS-1.7B's 27.41, but worse than the 4B-class Higgs Audio v3 (2.74) and Fish S2 Pro (4.40). The official documentation notes that these metrics only measure intelligibility and timbre similarity, not directly "realism."
- Includes a no-install Hugging Face Playground, and the SGLang Omni plugin provides an OpenAI-compatible /v1/audio/speech interface; licensed under Apache-2.0, and not previously covered in earlier issues.
Usage
- Command line: git clone Edge0-AI/Audio8_TTS, pip install -r requirements.txt, python audio8_tts_infer.py --model Edge0/Audio8-TTS-Realistic-0.6b --text "Text to synthesize" --output outputs/realistic.wav; for voice cloning, add --reference-audio reference.wav --reference-text "Transcript of the reference recording" (must exactly match the recording content).
- Transformers (version 4.57 up to but not including 5, with trust_remote_code=True): pass text, reference_audio, and reference_text to the processor, use temperature 0.8, top_p 0.95, and top_k 50 for generation, then decode to waveform using model.decode_audio.
- Long paragraphs should be split into short sentences: the reference audio, transcript, and target text share a context window of 2,048 positions.
Risks
- Quality varies significantly across languages: on CV3, the error rate is 48.574 for Thai and 32.279 for Finnish; on the MiniMax set, it is 31.349 for Cantonese; the support list does not imply equal usability across all languages.
- Official compatibility checks were performed only on H20 BF16; throughput and streaming performance have not been tested; issues such as mispronunciations, repetitions, omissions, or premature termination may occur, and the official documentation advises listening to each output before finalizing. This column has not tested the examples.
- Cloning someone else's voice requires authorization, and synthesized audio should be labeled as such when necessary.