RealTMS

Speech translation usually stops at text. RealTMS goes speech to speech across 15+ languages and adds a generated image of what was said, all from one Gradio interface.

• Speech-to-speech translation across 15+ language pairs
• Whisper handles speech recognition with good noise tolerance
• SeamlessM4T performs the multilingual neural machine translation
• Stable Diffusion generates a contextual image from the translated text
• Upload or record audio directly in the Gradio interface
• CPU-only fallback mode: a GPU speeds things up but is not required

• Four-stage pipeline: Whisper for speech recognition, SeamlessM4T for translation, gTTS for speech synthesis, and Stable Diffusion for image generation.
• Generated images hit 85% visual relevance against the translated text.