Sortformer Diarizer 4spk v1

Sortformer[1] is a novel end-to-end neural model for speaker diarization, trained with unconventional objectives compared to existing end-to-end diarization models.

Sortformer resolves permutation problem in diarization following the arrival-time order of the speech segments from each speaker.

Model Architecture

Sortformer consists of an L-size (18 layers) NeMo Encoder for Speech Tasks (NEST)[2] which is based on Fast-Conformer[3] encoder. Following that, an 18-layer Transformer[4] encoder with hidden size of 192, and two feedforward layers with 4 sigmoid outputs for each frame input at the top layer. More information can be found in the Sortformer paper[1].

NVIDIA NeMo

To train, fine-tune or perform diarization with Sortformer, you will need to install NVIDIA NeMo[5]. We recommend you install it after you've installed Cython and latest PyTorch version.

pip install git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[asr]

How to Use this Model

The model is available for use in the NeMo Framework[5], and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.

Loading the Model

from nemo.collections.asr.models import SortformerEncLabelModel

# load model from a downloaded file
diar_model = SortformerEncLabelModel.restore_from(restore_path="/path/to/diar_sortformer_4spk-v1.nemo", map_location=torch.device('cuda'), strict=False)
# load model from Hugging Face model card directly (You need a Hugging Face token)
diar_model = SortformerEncLabelModel.from_pretrained("nvidia/diar_sortformer_4spk-v1")

Input Format

Input to Sortformer can be an individual audio file:

audio_input="/path/to/multispeaker_audio1.wav"

or a list of paths to audio files:

audio_input=["/path/to/multispeaker_audio1.wav", "/path/to/multispeaker_audio2.wav"]

or a jsonl manifest file:

audio_input="/path/to/multispeaker_manifest.json"

where each line is a dictionary containing the following fields:

# Example of a line in `multispeaker_manifest.json`
{
    "audio_filepath": "/path/to/multispeaker_audio1.wav",  # path to the input audio file 
    "offset": 0, # offset (start) time of the input audio
    "duration": 600,  # duration of the audio, can be set to `null` if using NeMo main branch
}
{
    "audio_filepath": "/path/to/multispeaker_audio2.wav",  
    "offset": 900,
    "duration": 580,  
}

Getting Diarization Results

To perform speaker diarization and get a list of speaker-marked speech segments in the format 'begin_seconds, end_seconds, speaker_index', simply use:

predicted_segments = diar_model.diarize(audio=audio_input, batch_size=1)

To obtain tensors of speaker activity probabilities, use:

predicted_segments, predicted_probs = diar_model.diarize(audio=audio_input, batch_size=1, include_tensor_outputs=True)

Input

This model accepts single-channel (mono) audio sampled at 16,000 Hz.

The actual input tensor is a Ns x 1 matrix for each audio clip, where Ns is the number of samples in the time-series signal.
For instance, a 10-second audio clip sampled at 16,000 Hz (mono-channel WAV file) will form a 160,000 x 1 matrix.

Output

The output of the model is a T x S matrix, where:

S is the maximum number of speakers (in this model, S = 4).
T is the total number of frames, including zero-padding. Each frame corresponds to a segment of 0.08 seconds of audio.
Each element of the T x S matrix represents the speaker activity probability in the [0, 1] range. For example, a matrix element a(150, 2) = 0.95 indicates a 95% probability of activity for the second speaker during the time range [12.00, 12.08] seconds.

Train and evaluate Sortformer diarizer using NeMo

Training

Sortformer diarizer models are trained on 8 nodes of 8×NVIDIA Tesla V100 GPUs. We use 90 second long training samples and batch size of 4. The model can be trained using this example script and base config.

Evaluation

To evaluate Sortformer diarizer and save diarization results in RTTM format, use the inference example script:

python ${NEMO_GIT_FOLDER}/examples/speaker_tasks/diarization/neural_diarizer/e2e_diarize_speech.py 
 model_path="/path/to/diar_sortformer_4spk-v1.nemo" \
 manifest_filepath="/path/to/multispeaker_manifest_with_reference_rttms.json" \
 collar=COLLAR \
 out_rttm_dir="/path/to/output_rttms"

You can provide the post-processing YAML configs from post_processing folder to reproduce the optimized post-processing algorithm for each development dataset:

python ${NEMO_GIT_FOLDER}/examples/speaker_tasks/diarization/neural_diarizer/e2e_diarize_speech.py \
 model_path="/path/to/diar_sortformer_4spk-v1.nemo" \
 manifest_filepath="/path/to/multispeaker_manifest_with_reference_rttms.json" \
 collar=COLLAR \
 bypass_postprocessing=False \
 postprocessing_yaml="/path/to/postprocessing_config.yaml" \
 out_rttm_dir="/path/to/output_rttms"

Technical Limitations

The model operates in a non-streaming mode (offline mode).
It can detect a maximum of 4 speakers; performance degrades on recordings with 5 and more speakers.
The maximum duration of a test recording depends on available GPU memory. For an RTX A6000 48GB model, the limit is around 12 minutes.
The model was trained on publicly available speech datasets, primarily in English. As a result:
- Performance may degrade on non-English speech.
- Performance may also degrade on out-of-domain data, such as recordings in noisy conditions.

Datasets

Sortformer was trained on a combination of 2030 hours of real conversations and 5150 hours or simulated audio mixtures generated by NeMo speech data simulator[6]. All the datasets listed above are based on the same labeling method via RTTM format. A subset of RTTM files used for model training are processed for the speaker diarization model training purposes. Data collection methods vary across individual datasets. For example, the above datasets include phone calls, interviews, web videos, and audiobook recordings. Please refer to the Linguistic Data Consortium (LDC) website or dataset webpage for detailed data collection methods.

Training Datasets (Real conversations)

Fisher English (LDC)
2004-2010 NIST Speaker Recognition Evaluation (LDC)
Librispeech
AMI Meeting Corpus
VoxConverse-v0.3
ICSI
AISHELL-4
Third DIHARD Challenge Development (LDC)
2000 NIST Speaker Recognition Evaluation, split1 (LDC)

Training Datasets (Used to simulate audio mixtures)

2004-2010 NIST Speaker Recognition Evaluation (LDC)
Librispeech

Performance

Evaluation dataset specifications

Dataset	DIHARD3-Eval	CALLHOME-part2	CALLHOME-part2	CALLHOME-part2	CH109
Number of Speakers	≤ 4 speakers	2 speakers	3 speakers	4 speakers	2 speakers
Collar (sec)	0.0s	0.25s	0.25s	0.25s	0.25s
Mean Audio Duration (sec)	453.0s	73.0s	135.7s	329.8s	552.9s

Diarization Error Rate (DER)

All evaluations include overlapping speech.
Bolded and italicized numbers represent the best-performing Sortformer evaluations.
Post-Processing (PP) is optimized on two different held-out dataset splits.
- YAML file for DH3-dev Optimized Post-Processing
- YAML file for CallHome-part1 Optimized Post-Processing

Dataset	DIHARD3-Eval	CALLHOME-part2	CALLHOME-part2	CALLHOME-part2	CH109
DER diar_sortformer_4spk-v1	16.28	6.49	10.01	14.14	*6.27*
DER diar_sortformer_4spk-v1 + DH3-dev Opt. PP	*14.76*	-	-	-	-
DER diar_sortformer_4spk-v1 + CallHome-part1 Opt. PP	-	*5.85*	*8.46*	*12.59*	6.86

Real Time Factor (RTFx)

All tests were measured on RTX A6000 48GB with batch size of 1. Post-processing is not included in RTFx calculations.

Datasets	DIHARD3-Eval	CALLHOME-part2	CALLHOME-part2	CALLHOME-part2	CH109
RTFx diar_sortformer_4spk-v1	437	1053	915	545	415

NVIDIA Riva: Deployment

NVIDIA Riva, is an accelerated speech AI SDK deployable on-prem, in all clouds, multi-cloud, hybrid, on edge, and embedded. Additionally, Riva provides:

World-class out-of-the-box accuracy for the most common languages with model checkpoints trained on proprietary data with hundreds of thousands of GPU-compute hours
Best in class accuracy with run-time word boosting (e.g., brand and product names) and customization of acoustic model, language model, and inverse text normalization
Streaming speech recognition, Kubernetes compatible scaling, and enterprise-grade support

Although this model isn’t supported yet by Riva, the list of supported models is here.
Check out Riva live demo.

References

[1] Sortformer: Seamless Integration of Speaker Diarization and ASR by Bridging Timestamps and Tokens

[2] NEST: Self-supervised Fast Conformer as All-purpose Seasoning to Speech Processing Tasks

[3] Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition

[4] Attention is all you need

[5] NVIDIA NeMo Framework

[6] NeMo speech data simulator

Licence

License to use this model is covered by the CC-BY-NC-4.0. By downloading the public and release version of the model, you accept the terms and conditions of the CC-BY-NC-4.0 license.

nvidia
/

diar_sortformer_4spk-v1