---
library_name: nemo
---
# CHiME8 DASR NeMo Baseline Models

- The model files in this repository are the models used in this paper [The CHiME-7 Challenge: System Description and Performance of
NeMo Team’s DASR System](https://arxiv.org/pdf/2310.12378.pdf).
- These models are needed to execute the CHiME8-DASR baseline [CHiME8-DASR-Baseline NeMo](https://github.com/chimechallenge/C8DASR-Baseline-NeMo/tree/main/scripts/chime8)
- VAD, Diarization and ASR models are all based on [NVIDIA NeMo Conversational AI Toolkits](https://github.com/NVIDIA/NeMo).

## 1. Voice Activity Detection (VAD) Model: 
### **[**MarbleNet_frame_VAD_chime7_Acrobat.nemo**](https://huggingface.co/chime-dasr/nemo_baseline_models/blob/main/MarbleNet_frame_VAD_chime7_Acrobat.nemo)**   
- This model is based on [NeMo MarbleNet VAD model](https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/speech_classification/models.html#marblenet-vad).
- For validation, we use dataset comprises the CHiME-6 development subset as well as 50 hours of simulated audio data.
- The simulated data is generated using the [NeMo multi-speaker data simulator](https://github.com/NVIDIA/NeMo/blob/main/tutorials/tools/Multispeaker_Simulator.ipynb)
on [VoxCeleb1&2 datasets](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/vox1.html)
- The multi-speaker data simulation results in a total of 2,000 hours of audio, of which approximately 30% is silence.
- The Model training incorporates [SpecAugment](https://arxiv.org/abs/1904.08779) and noise augmentation through [MUSAN noise dataset](https://arxiv.org/abs/1510.08484).


## 2. Speaker Diarization Model: Multi-scale Diarization Decoder (MSDD-v2)
### **[**MSDD_v2_PALO_100ms_intrpl_3scales.nemo**](https://huggingface.co/chime-dasr/nemo_baseline_models/blob/main/MSDD_v2_PALO_100ms_intrpl_3scales.nemo)**

 Our DASR system is based on the speaker diarization system using the multi-scale diarization decoder (MSDD).
  - MSDD Reference: [Park et al. (2022)](https://arxiv.org/pdf/2203.15974.pdf)
- MSDD-v2 speaker diarization system employs a multi-scale embedding approach and utilizes TitaNet speaker embedding extractor.
  - TitaNet Reference: [Koluguri et al. (2022)](https://arxiv.org/abs/2110.04410)
  - TitaNet Model is included in [MSDD-v2 .nemo checkpoint file](https://huggingface.co/chime-dasr/nemo_baseline_models/blob/main/MSDD_v2_PALO_100ms_intrpl_3scales.nemo).
- Unlike the system that uses a multi-layer LSTM architecture, we employ a four-layer Transformer architecture with a hidden size of 384.
- This neural model generates logit values indicating speaker existence.
- Our diarization model is trained on approximately 3,000 hours of simulated audio mixture data from the same multi-speaker data simulator used in VAD model training, drawing from VoxCeleb1&2 and LibriSpeech datasets.
  - LibriSpeech Reference: [OpenSLR Download](https://www.openslr.org/12),[LibriSpeech, Panayotov et al. (2015)](https://ieeexplore.ieee.org/document/7178964)
- MUSAN noise is also used for adding additive background noise, focusing on music and broadband noise.


## 3. Automatic Speech Recognition (ASR) model
### **[**FastConformerXL-RNNT-chime7-GSS-finetuned.nemo**](https://huggingface.co/chime-dasr/nemo_baseline_models/blob/main/FastConformerXL-RNNT-chime7-GSS-finetuned.nemo)**
- This ASR model is based on [NeMo FastConformer XL model](https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/models.html#fast-conformer). 
- Single-channel audio generated using a multi-channel front-end (Guided Source Separation, GSS) is transcribed using a 0.6B parameter Conformer-based transducer (RNNT) model.
  - Model Reference: [Gulati et al. (2020)](https://arxiv.org/abs/2005.08100)
- The model was initialized using a publicly available NeMo checkpoint.
  - NeMo Checkpoint: [NGC Model Card: Conformer Transducer XL](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_en_conformer_transducer_xlarge)
- This model was then fine-tuned on the CHiME-7 train and dev set, which includes the CHiME-6 and Mixer6 training subsets, after processing the data through the multi-channel ASR front-end, utilizing ground-truth diarization.
  - Fine-Tuning Details:
    - Fine-tuning Duration: 35,000 updates
    - Batch Size: 128


## 4. Language Model for ASR Decoding: KenLM Model
### **[**ASR_LM_chime7_only.kenlm**](https://huggingface.co/chime-dasr/nemo_baseline_models/blob/main/ASR_LM_chime7_only.kenlm)**

- This KenLM model is trained solely on CHiME7-DASR datasets (Mixer6, CHiME6, DipCo).
- We apply a word-piece level N-gram language model using byte-pair-encoding (BPE) tokens.
- This approach utilizes the SentencePiece and KenLM toolkits, based on the transcription of CHiME-7 train and dev sets.
  - SentencePiece: [Kudo and Richardson (2018)](https://arxiv.org/abs/1808.06226)
  - KenLM: [KenLM GitRepo](https://github.com/kpu/kenlm)
- The token sets of our ASR and LM models were matched to ensure consistency.
- To combine several N-gram models with equal weights, we used the OpenGrm library.
  - OpenGrm: [Roark et al. (2012)](https://aclanthology.org/P12-3011/)
- MAES decoding was employed for the transducer, which accelerates the decoding process.
  - MAES Decoding: [Kim et al. (2020)](https://ieeexplore.ieee.org/document/9250505)
- As expected, integrating the beam-search decoder with the language model significantly enhances the performance of the end-to-end model compared to its pure counterpart.