Can NVIDIA NeMo preserve voice identity while reducing a foreign accent in prerecorded speech?

Hello everyone,

I’m a software engineer building an AI-powered application and would really appreciate guidance from the NVIDIA NeMo or Speech AI team.

Our goal is not text-to-speech, voice cloning, dubbing, or translation.

Instead, we are trying to build an AI pipeline that preserves the speaker completely while improving only the pronunciation.

The intended workflow is:

  1. A user uploads a prerecorded video.
  2. Our application extracts the original audio.
  3. If necessary, we also provide an accurate transcript to help the AI correctly interpret words that may be difficult to recognize because of a strong foreign accent.
  4. The AI processes the speech while preserving:
    • the speaker’s original voice identity,
    • emotions,
    • speaking style,
    • pacing,
    • pauses,
    • timing,
    • and overall delivery.
  5. The only desired change is improving pronunciation and reducing the foreign accent.
  6. The processed audio is merged back into the original video.

The output should sound like the same person, with the same timing and emotions, but with a significantly more natural pronunciation.

During the past few weeks, I’ve evaluated several commercial speech AI solutions (Krisp, Sanas, Deepgram, Speechmatics, Rime, and others), but I haven’t found a solution that clearly supports this exact workflow.

My questions are:

  1. Is this currently possible using NVIDIA NeMo, Riva, NIM, or another NVIDIA speech technology?
  2. Would this require combining multiple models, or is there an existing speech-to-speech pipeline that could achieve this?
  3. Are there any NVIDIA reference architectures or research projects addressing this use case?
  4. If not currently supported, is this an area that NVIDIA is actively researching?

Any technical guidance or recommendations would be greatly appreciated.

Thank you very much!