Halo NeuroAI Releases Sopro V2 Turbo, a 120M-Parameter Open-Source Voice Cloning TTS with Native European Portuguese Support

Halo NeuroAI has announced its new TTS model family, Sopro V2, and released the fastest variant, sopro-v2-turbo, as open source. This 120M-parameter voice cloning (text-to-speech) model is the first open TTS model to natively support European Portuguese, in addition to English, German, and French.

In terms of performance, on an Apple M3 CPU, it achieves 0.24 RTF (real-time factor) for offline generation and approximately 300ms time-to-first-audio for streaming, allowing smooth execution on laptop CPUs or browsers. On an H100 GPU, it reaches 0.07 RTF for offline generation and approximately 200ms time-to-first-audio for streaming.

This project was initiated to provide European Portuguese voices to patients who have lost their voices due to conditions such as ALS. Existing commercial APIs (OpenAI, ElevenLabs, etc.) suffered from inaccuracies in European Portuguese pronunciation or long latency due to bias toward Brazilian Portuguese data, making them difficult to use in production. To address these issues, Sopro V2 optimizes a small model architecture, starting from Sesame's CSM architecture and improving components with a SentencePiece tokenizer, a Flow-matching acoustic head, and a causal vocoder.