Important

NeMo 2.0 is an experimental feature and currently released in the dev container only: nvcr.io/nvidia/nemo:dev. Please refer to NeMo 2.0 overview for information on getting started.

Speech AI Models

NVIDIA NeMo Framework supports the training and customization of Speech AI models, specifically designed to enable voice-based interfaces for conversational AI applications. A range of speech tasks are supported, including Automatic Speech Recognition (ASR), Speaker Diarization, and Text-to-Speech (TTS), which we highlight below.

Automatic Speech Recognition (ASR)

Automatic Speech Recognition is the task of generating transcriptions of what was spoken in an audio file.

Latest ASR Models Developed by the NVIDIA NeMo Team

Model family	Decoder type	Useful links
Canary	AED (Attention-based Encoder-Decoder)	Docs, Paper, HF space
Parakeet	CTC, RNN-T, TDT, TDT-CTC hybrid	Docs, HF space

Key features of NeMo ASR include:

• Pretrained ASR models, many topping the HuggingFace Open ASR Leaderboard

• Model checkpoints specialized for real-time speech recognition

• LM decoding

• Keyword spotting

Find more details in the Developer Docs.

Speaker Diarization

Speaker diarization is the process of partitioning an audio stream into segments based on the identity of each speaker. Essentially, it answers the question, “Who spoke when?”

Latest Speaker Diarization Models Developed by the NVIDIA NeMo Team

Model name	Useful links
MSDD (Multiscale Diarization Decoder)	Docs, Paper, HF space

Find more details in the Developer Docs.

Text-To-Speech (TTS)

Text-to-Speech is a technology that converts textual inputs into natural human speech.

Latest TTS Models Developed by the NVIDIA NeMo Team

Model name	Useful links
T5-TTS	Paper, Blog post

Find more details in the Developer Docs.

Speech AI Tools

NeMo Framework also includes a large set of Speech AI tools for dataset preparation, model evaluation, and text normalization.