For local speech-to-text, Whisper remains the gold standard - you can run it locally with good accuracy across languages. For speech-to-speech, you'd typically chain Whisper with a local TTS model like Coqui TTS or use something like Tortoise TTS for higher quality but slower processing. The key is balancing accuracy, speed, and resource usage based on your specific use case. If you're doing content creation workflows, consider what post-processing you might need - sometimes the raw transcription needs structure and enhancement beyond just accurate words.
+1 on the post-processing point. Raw Whisper output is ~90% there but punctuation, grammar, and formatting are the missing piece.
I built MumbleFlow to address exactly this — whisper.cpp for STT plus llama.cpp for smart text cleanup, all running on-device. Metal/CUDA accelerated, sub-second latency on Apple Silicon. Global hotkey works in any app.
Comments
For local speech-to-text, Whisper remains the gold standard - you can run it locally with good accuracy across languages. For speech-to-speech, you'd typically chain Whisper with a local TTS model like Coqui TTS or use something like Tortoise TTS for higher quality but slower processing. The key is balancing accuracy, speed, and resource usage based on your specific use case. If you're doing content creation workflows, consider what post-processing you might need - sometimes the raw transcription needs structure and enhancement beyond just accurate words.
+1 on the post-processing point. Raw Whisper output is ~90% there but punctuation, grammar, and formatting are the missing piece.
I built MumbleFlow to address exactly this — whisper.cpp for STT plus llama.cpp for smart text cleanup, all running on-device. Metal/CUDA accelerated, sub-second latency on Apple Silicon. Global hotkey works in any app.
$5 one-time, no cloud, no subscription. https://mumble.helix-co.com