Build Pipelines
Real-Time Voice Translation
Whisper + MarianMT + Edge-TTS delivering speech-to-speech translation in 1.3 seconds across four languages.
The problem
Speech-to-speech translation needs to chain transcription, translation, and synthesis models without breaking conversational flow.
- The translation is accurate but arrives so late the conversation has already moved on.
- Chaining three models together usually means three separate points of failure.
- A natural pause in speech gets mistaken for the end of a sentence, and the translation is wrong.
What NEO built
NEO chained Whisper for transcription, MarianMT for translation, and Edge-TTS for synthesis, using streaming inference and audio chunking to keep the pipeline real-time.
The result
1.3s end-to-end latency
Delivers full speech-to-speech translation across four languages in 1.3s end-to-end.
From the blog · 8 min
Real-Time Voice Translation Pipeline with Sub-2-Second Latency
How NEO built a production voice translation system using Whisper, MarianMT, and Edge-TTS that delivers end-to-end speech-to-speech translation in 1.3 seconds across English, Spanish, French, and German.
Try this in your workspace
Paste this into NEO chat to kick off the same workflow on your own data.
Chain speech-to-text, translation, and text-to-speech into a real-time pipeline so I can speak in one language and have it come out in another with minimal lag.
Paste it in · review the plan · get the diff
Get NEO
