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.
Streaming Sortformer[2] employs an Arrival-Order Speaker Cache (AOSC) to store frame-level acoustic embeddings of previously observed speakers.
Sortformer resolves permutation problem in diarization following the arrival-time order of the speech segments from each speaker.
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.
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.
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].
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.
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
python
1from nemo.collections.asr.models import SortformerEncLabelModel
23# load model from Hugging Face model card directly (You need a Hugging Face token)4diar_model = SortformerEncLabelModel.from_pretrained("nvidia/diar_streaming_sortformer_4spk-v2")56# If you have a downloaded model in "/path/to/diar_streaming_sortformer_4spk-v2.nemo", load model from a downloaded file7diar_model = SortformerEncLabelModel.restore_from(restore_path="/path/to/diar_streaming_sortformer_4spk-v2.nemo", map_location='cuda', strict=False)89# switch to inference mode10diar_model.eval()
Input Format
Input to Sortformer can be an individual audio file:
1import numpy as np
2audio_input = np.random.randn(16000*10).astype(np.float32)# 10 sec at 16kHz3# or a list of arrays4audio_input =[audio_array1, audio_array2]5diar_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:
yaml
1# Example of a line in `multispeaker_manifest.json`2{3"audio_filepath":"/path/to/multispeaker_audio1.wav",# path to the input audio file 4"offset":0,# offset (start) time of the input audio5"duration":600,# duration of the audio, can be set to `null` if using NeMo main branch6}7{8"audio_filepath":"/path/to/multispeaker_audio2.wav",9"offset":900,10"duration":580,11}
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:
Configuration
Latency
RTF
CHUNK_SIZE
RIGHT_CONTEXT
FIFO_SIZE
UPDATE_PERIOD
SPEAKER_CACHE_SIZE
very high latency
30.4s
0.002
340
40
40
300
188
low latency
1.04s
0.093
6
7
188
144
188
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.
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.
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 approximately 5,000 hours of audio, combining real conversations and simulated audio mixtures generated using the NeMo speech data simulator[7].
All datasets used in training follow the RTTM labeling format. A subset of the RTTM files were processed specifically for speaker diarization model training.
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.
