語音模型與邊緣推論
Sopro V2 Turbo Releases a 120M-Parameter Speech Model with Two-Step Acoustic Solving for Streaming Synthesis on CPUs
Sopro V2 Turbo distills a 500M-parameter teacher into a 120M-parameter model, then uses reflow to reduce acoustic solving from 32 steps to two. The model can produce its first audio chunk in about 300 milliseconds on an M3 CPU and supports in-browser inference, but it currently covers only four European languages and does not include training code.
Halo Research has released Sopro V2 Turbo’s Apache 2.0-licensed weights, PyTorch inference code, and an ONNX browser version. The model supports English, German, French, and European Portuguese, and can perform zero-shot voice cloning from 5–20 seconds of reference audio. On an Apple M3 CPU, the company reports an offline real-time factor of 0.24 and a streaming real-time factor of 0.21, with the first audio chunk arriving in about 300 milliseconds. This means 10 seconds of speech could theoretically be generated in roughly two seconds, but the figures come from a single stream under default settings and cannot be directly extrapolated to high-concurrency services.
The architecture first uses an autoregressive decoder with RoPE, RMSNorm, and grouped-query attention to produce semantic tokens. A flow-matching acoustic head then generates a mel spectrogram, which is passed to a causal Vocos variant with a three-frame lookahead for streaming audio output. The speech tokenizer is warm-started from the Whisper large-v3 encoder and uses an FSQ bottleneck before being distilled into a 20M-parameter version. It does not require a transcript of the reference speech, so edge deployments do not need to run a separate ASR model.
The training pipeline begins with a 500M-parameter teacher model. After three rounds of DPO, it is distilled into a 120M-parameter student model. Reflow-based self-distillation then reduces the acoustic solver from 32 steps to two. On Seed-TTS test-en, the company reports a WER of 1.51 and speaker similarity of 0.644 for the streaming version; on LibriSpeech, it reports a WER of 1.88. These results were scored automatically using Whisper large-v3 and WavLM. Some baseline figures were taken directly from other papers, so they do not constitute an independent comparison conducted with uniform hardware and code.
The engineering limitations are equally clear: pronunciation remains unreliable for mixed-language text, numbers, and symbols, and streaming quality does not fully match the offline version. Although the weights and inference code are public, the training code will not be released in the near term, and the model includes no watermarking. Chinese-language teams can draw on its small-vocabulary design, ASR-aligned tokenizer, and two-step acoustic distillation approach, but the existing checkpoint cannot synthesize Chinese directly. The next key step is for third parties to reproduce the latency results and evaluate subjective audio quality and voice-cloning safety.