diar_streaming_sortformer_4spk-v2

A new version of streaming Sortformer v2.1 has been released, providing greater robustness for meeting speech.

Streaming Sortformer Diarizer 4spk v2

<style> img { display: inline; } </style>

|

<!-- | [](#datasets) -->

This model is a streaming version of Sortformer diarizer. 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.

<div align="center"> <img src="figures/sortformer_intro.png" width="750" /> </div>

Streaming Sortformer[2] employs an Arrival-Order Speaker Cache (AOSC) to store frame-level acoustic embeddings of previously observed speakers.

<div align="center"> <img src="figures/aosc_3spk_example.gif" width="1400" /> </div> <div align="center"> <img src="figures/aosc_4spk_example.gif" width="1400" /> </div>

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

This speaker diarization model can be used to enable the NeMo Voice Agent to recognize speakers in conversations. See the NeMo Voice Agent and the YAML configuration for more details.

Discover more from NVIDIA:

For documentation, deployment guides, enterprise-ready APIs, and the latest open models—including Nemotron and other cutting-edge speech, translation, and generative AI—visit the NVIDIA Developer Portal at developer.nvidia.com. Join the community to access tools, support, and resources to accelerate your development with NVIDIA’s NeMo, Riva, NIM, and foundation models.<br>

Explore more from NVIDIA: <br>

What is Nemotron?<br> NVIDIA Developer Nemotron<br> NVIDIA Riva Speech<br> NeMo Documentation<br>

Model Architecture

Streaming sortformer employs pre-encode layer in the Fast-Conformer to generate speaker-cache. At each step, speaker cache is filtered to only retain the high-quality speaker cache vectors.

<div align="center"> <img src="figures/streaming_steps.png" width="1400" /> </div>

Aside from speaker-cache management part, streaming Sortformer follows the architecture of the offline version of Sortformer. Sortformer consists of an L-size (17 layers) NeMo Encoder for Speech Tasks (NEST)[3] which is based on Fast-Conformer[4] encoder. Following that, an 18-layer Transformer[5] 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 Streaming Sortformer paper[2].

<div align="center"> <img src="figures/sortformer-v1-model.png" width="450" /> </div>

NVIDIA NeMo

To train, fine-tune or perform diarization with Sortformer, you will need to install NVIDIA NeMo[6]. We recommend you install it after you've installed Cython and latest PyTorch version.

apt-get update && apt-get install -y libsndfile1 ffmpeg
pip install Cython packaging
pip install 'git+https://github.com/NVIDIA/NeMo.git@main#egg=nemo_toolkit[asr]'

Discover more from NVIDIA:

For documentation, deployment guides, enterprise-ready APIs, and the latest open models—including Nemotron and other cutting-edge speech, translation, and generative AI—visit the NVIDIA Developer Portal at developer.nvidia.com. Join the community to access tools, support, and resources to accelerate your development with NVIDIA’s NeMo, Riva, NIM, and foundation models.<br>

Explore more from NVIDIA: <br>

What is Nemotron?<br> NVIDIA Developer Nemotron<br> NVIDIA Riva Speech<br> NeMo Documentation<br>

🚀 Quick Start: Run Diarization Now

Here is a short example script that loads the model, runs diarization on a WAV file, and prints the results:

from nemo.collections.asr.models import SortformerEncLabelModel
diar_model = SortformerEncLabelModel.from_pretrained("nvidia/diar_streaming_sortformer_4spk-v2")
diar_model.eval()

diar_model.sortformer_modules.chunk_len = 340
diar_model.sortformer_modules.chunk_right_context = 40
diar_model.sortformer_modules.fifo_len = 40
diar_model.sortformer_modules.spkcache_update_period = 300

predicted_segments = diar_model.diarize(audio=["/path/to/your/audio.wav"], batch_size=1)

for segment in predicted_segments[0]:
    print(segment)

How to Use this Model

The model is available for use in the NeMo Framework[6], 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 Hugging Face model card directly (You need a Hugging Face token)
diar_model = SortformerEncLabelModel.from_pretrained("nvidia/diar_streaming_sortformer_4spk-v2")

# If you have a downloaded model in "/path/to/diar_streaming_sortformer_4spk-v2.nemo", load model from a downloaded file
diar_model = SortformerEncLabelModel.restore_from(restore_path="/path/to/diar_streaming_sortformer_4spk-v2.nemo", map_location='cuda', strict=False)

# switch to inference mode
diar_model.eval()

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 numpy array (single or list):

import numpy as np
audio_input = np.random.randn(16000 * 10).astype(np.float32)  # 10 sec at 16kHz
# or a list of arrays
audio_input = [audio_array1, audio_array2]
diar_model.diarize(audio=audio_input, batch_size=2, sample_rate=16000)

Note: When using numpy arrays, you MUST specify a correct sample_rate in diar_model.diarize() function. Default sample_rate is 16000.

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,  
}

Setting up Streaming Configuration

Streaming configuration is defined by the following parameters, all measured in 80ms frames:

• CHUNK_SIZE: The number of frames in a processing chunk.
• RIGHT_CONTEXT: The number of future frames attached after the chunk.
• FIFO_SIZE: The number of previous frames attached before the chunk, from the FIFO queue.
• UPDATE_PERIOD: The number of frames extracted from the FIFO queue to update the speaker cache.
• SPEAKER_CACHE_SIZE: The total number of frames in the speaker cache.

Here are recommended configurations for different scenarios:

For clarity on the metrics used in the table:

• Latency: Refers to Input Buffer Latency, calculated as CHUNK_SIZE + RIGHT_CONTEXT. This value does not include computational processing time.
• Real-Time Factor (RTF): Characterizes processing speed, calculated as the time taken to process an audio file divided by its duration. RTF values are measured with a batch size of 1 on an NVIDIA RTX 6000 Ada Generation GPU.

To set streaming configuration, use:

diar_model.sortformer_modules.chunk_len = CHUNK_SIZE
diar_model.sortformer_modules.chunk_right_context = RIGHT_CONTEXT
diar_model.sortformer_modules.fifo_len = FIFO_SIZE
diar_model.sortformer_modules.spkcache_update_period = UPDATE_PERIOD
diar_model.sortformer_modules.spkcache_len = SPEAKER_CACHE_SIZE
diar_model.sortformer_modules._check_streaming_parameters()

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)

Note that if you are feeding a list of numpy arrays, you MUST provide the sample_rate in integer format.

predicted_segments, predicted_probs = diar_model.diarize(audio=[np_array1, np_array2], batch_size=2, sample_rate=16000)

🔬 For more detailed evaluations (DER)

If you need to perform a comprehensive evaluation and calculate the Diarization Error Rate (DER) across different parameter settings, use the NeMo example script e2e_diarize_speech.py. This script allows you to test the streaming behavior of the model by adjusting key parameters like chunk_len, fifo_len, and spkcache_update_period.

python ${NEMO_ROOT}/examples/speaker_tasks/diarization/neural_diarizer/e2e_diarize_speech.py \
    model_path="/path/to/diar_sortformer_4spk_v1.nemo" \
    dataset_manifest="/path/to/diarization_manifest.json" \
    batch_size=1 \
    spkcache_len=188 \
    spkcache_update_period=300 \
    fifo_len=40 \
    chunk_len=340 \
    chunk_right_context=40

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 an 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.

Inference

Sortformer diarizer models can be performed with post-processing algorithms using inference example script. If you provide the post-processing YAML configs in post_processing folder to reproduce the optimized post-processing algorithm for each development dataset.

Technical Limitations

• The model operates in a streaming mode (online mode).
• It can detect a maximum of 4 speakers; performance degrades on recordings with 5 and more speakers.
• While the model is designed for long-form audio and can handle recordings that are several hours long, performance may degrade on very long recordings.
• 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 2445 hours of real conversations and 5150 hours or simulated audio mixtures generated by NeMo speech data simulator[7]. 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)
• AMI Meeting Corpus
• VoxConverse-v0.3
• ICSI
• AISHELL-4
• Third DIHARD Challenge Development (LDC)
• 2000 NIST Speaker Recognition Evaluation, split1 (LDC)
• DiPCo
• AliMeeting

Training Datasets (Used to simulate audio mixtures)

• 2004-2010 NIST Speaker Recognition Evaluation (LDC)
• Librispeech

Performance

Evaluation data specifications

Diarization Error Rate (DER)

• All evaluations include overlapping speech.
• Collar tolerance is 0s for DIHARD III Eval, and 0.25s for CALLHOME-part2 and CH109.
• Post-Processing (PP) is optimized on two different held-out dataset splits.
  • DIHARD III Dev Optimized Post-Processing for DIHARD III Eval
  • CALLHOME-part1 Optimized Post-Processing for CALLHOME-part2 and CH109

NVIDIA Riva: Deployment

Streaming Sortformer is deployed via NVIDIA RIVA ASR - Speech Recognition with Speaker Diarization 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 For more information on NVIDIA RIVA, see the list of supported models is here.
  Also check out the Riva live demo.

References

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

[2] Streaming Sortformer: Speaker Cache-Based Online Speaker Diarization with Arrival-Time Ordering

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

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

[5] Attention is all you need

[6] NVIDIA NeMo Framework

[7] NeMo speech data simulator

Licence

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