- Create environment:
conda create -n diart python=3.8
conda activate diart
- Install
PortAudio
andsoundfile
:
conda install portaudio
conda install pysoundfile -c conda-forge
-
Install pyannote.audio
pip install pyannote.audio==2.0.1
Note: starting from version 0.4, installing pyannote.audio is mandatory to run the default system or to use pyannote-based models. In any other case, this step can be ignored.
- Install diart:
pip install diart
A recorded conversation:
diart.stream /path/to/audio.wav
A live conversation:
diart.stream microphone
See diart.stream -h
for more options.
Use RealTimeInference
to easily run a pipeline on an audio source and write the results to disk:
from diart.sources import MicrophoneAudioSource
from diart.inference import RealTimeInference
from diart.pipelines import OnlineSpeakerDiarization
from diart.sinks import RTTMWriter
pipeline = OnlineSpeakerDiarization()
mic = MicrophoneAudioSource(pipeline.config.sample_rate)
inference = RealTimeInference(pipeline, mic, do_plot=True)
inference.attach_observers(RTTMWriter("/output/file.rttm"))
inference()
For inference and evaluation on a dataset we recommend to use Benchmark
(see notes on reproducibility).
Third-party models can be integrated seamlessly by subclassing SegmentationModel
and EmbeddingModel
:
import torch
from typing import Optional
from diart.models import EmbeddingModel
from diart.pipelines import PipelineConfig, OnlineSpeakerDiarization
from diart.sources import MicrophoneAudioSource
from diart.inference import RealTimeInference
class MyEmbeddingModel(EmbeddingModel):
def __init__(self):
super().__init__()
self.my_pretrained_model = load("my_model.ckpt")
def __call__(
self,
waveform: torch.Tensor,
weights: Optional[torch.Tensor] = None
) -> torch.Tensor:
return self.my_pretrained_model(waveform, weights)
config = PipelineConfig(embedding=MyEmbeddingModel())
pipeline = OnlineSpeakerDiarization(config)
mic = MicrophoneAudioSource(config.sample_rate)
inference = RealTimeInference(pipeline, mic)
inference()
Diart implements a hyper-parameter optimizer based on optuna that allows you to tune any pipeline to any dataset.
diart.tune /wav/dir --reference /rttm/dir --output /output/dir
See diart.tune -h
for more options.
from diart.optim import Optimizer
optimizer = Optimizer("/wav/dir", "/rttm/dir", "/output/dir")
optimizer(num_iter=100)
This will write results to an sqlite database in /output/dir
.
For bigger datasets, it is sometimes more convenient to run multiple optimization processes in parallel. To do this, create a study on a recommended DBMS (e.g. MySQL or PostgreSQL) making sure that the study and database names match:
mysql -u root -e "CREATE DATABASE IF NOT EXISTS example"
optuna create-study --study-name "example" --storage "mysql://root@localhost/example"
You can now run multiple identical optimizers pointing to this database:
diart.tune /wav/dir --reference /rttm/dir --storage mysql://root@localhost/example
or in python:
from diart.optim import Optimizer
from optuna.samplers import TPESampler
import optuna
db = "mysql://root@localhost/example"
study = optuna.load_study("example", db, TPESampler())
optimizer = Optimizer("/wav/dir", "/rttm/dir", study)
optimizer(num_iter=100)
For a more advanced usage, diart also provides building blocks that can be combined to create your own pipeline.
Streaming is powered by RxPY, but the blocks
module is completely independent and can be used separately.
Obtain overlap-aware speaker embeddings from a microphone stream:
import rx.operators as ops
import diart.operators as dops
from diart.sources import MicrophoneAudioSource
from diart.blocks import SpeakerSegmentation, OverlapAwareSpeakerEmbedding
segmentation = SpeakerSegmentation.from_pyannote("pyannote/segmentation")
embedding = OverlapAwareSpeakerEmbedding.from_pyannote("pyannote/embedding")
sample_rate = segmentation.model.get_sample_rate()
mic = MicrophoneAudioSource(sample_rate)
stream = mic.stream.pipe(
# Reformat stream to 5s duration and 500ms shift
dops.regularize_audio_stream(sample_rate),
ops.map(lambda wav: (wav, segmentation(wav))),
ops.starmap(embedding)
).subscribe(on_next=lambda emb: print(emb.shape))
mic.read()
Output:
torch.Size([4, 512])
torch.Size([4, 512])
torch.Size([4, 512])
...
Diart is also compatible with the WebSocket protocol to serve pipelines on the web.
In the following example we build a minimal server that receives audio chunks and sends back predictions in RTTM format:
from diart.pipelines import OnlineSpeakerDiarization
from diart.sources import WebSocketAudioSource
from diart.inference import RealTimeInference
pipeline = OnlineSpeakerDiarization()
source = WebSocketAudioSource(pipeline.config.sample_rate, "localhost", 7007)
inference = RealTimeInference(pipeline, source, do_plot=True)
inference.attach_hooks(lambda ann_wav: source.send(ann_wav[0].to_rttm()))
inference()
Diart is the official implementation of the paper Overlap-aware low-latency online speaker diarization based on end-to-end local segmentation by Juan Manuel Coria, Hervé Bredin, Sahar Ghannay and Sophie Rosset.
We propose to address online speaker diarization as a combination of incremental clustering and local diarization applied to a rolling buffer updated every 500ms. Every single step of the proposed pipeline is designed to take full advantage of the strong ability of a recently proposed end-to-end overlap-aware segmentation to detect and separate overlapping speakers. In particular, we propose a modified version of the statistics pooling layer (initially introduced in the x-vector architecture) to give less weight to frames where the segmentation model predicts simultaneous speakers. Furthermore, we derive cannot-link constraints from the initial segmentation step to prevent two local speakers from being wrongfully merged during the incremental clustering step. Finally, we show how the latency of the proposed approach can be adjusted between 500ms and 5s to match the requirements of a particular use case, and we provide a systematic analysis of the influence of latency on the overall performance (on AMI, DIHARD and VoxConverse).
If you found diart useful, please make sure to cite our paper:
@inproceedings{diart,
author={Coria, Juan M. and Bredin, Hervé and Ghannay, Sahar and Rosset, Sophie},
booktitle={2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
title={Overlap-Aware Low-Latency Online Speaker Diarization Based on End-to-End Local Segmentation},
year={2021},
pages={1139-1146},
doi={10.1109/ASRU51503.2021.9688044},
}
Diart aims to be lightweight and capable of real-time streaming in practical scenarios. Its performance is very close to what is reported in the paper (and sometimes even a bit better).
To obtain the best results, make sure to use the following hyper-parameters:
Dataset | latency | tau | rho | delta |
---|---|---|---|---|
DIHARD III | any | 0.555 | 0.422 | 1.517 |
AMI | any | 0.507 | 0.006 | 1.057 |
VoxConverse | any | 0.576 | 0.915 | 0.648 |
DIHARD II | 1s | 0.619 | 0.326 | 0.997 |
DIHARD II | 5s | 0.555 | 0.422 | 1.517 |
diart.benchmark
and diart.inference.Benchmark
can run, evaluate and measure the real-time latency of the pipeline. For instance, for a DIHARD III configuration:
diart.benchmark /wav/dir --reference /rttm/dir --tau=0.555 --rho=0.422 --delta=1.517 --segmentation pyannote/segmentation@Interspeech2021
or using the inference API:
from diart.inference import Benchmark
from diart.pipelines import OnlineSpeakerDiarization, PipelineConfig
from diart.models import SegmentationModel
config = PipelineConfig(
# Set the model used in the paper
segmentation=SegmentationModel.from_pyannote("pyannote/segmentation@Interspeech2021"),
step=0.5,
latency=0.5,
tau_active=0.555,
rho_update=0.422,
delta_new=1.517
)
pipeline = OnlineSpeakerDiarization(config)
benchmark = Benchmark("/wav/dir", "/rttm/dir")
benchmark(pipeline)
This pre-calculates model outputs in batches, so it runs a lot faster.
See diart.benchmark -h
for more options.
For convenience and to facilitate future comparisons, we also provide the expected outputs of the paper implementation in RTTM format for every entry of Table 1 and Figure 5. This includes the VBx offline topline as well as our proposed online approach with latencies 500ms, 1s, 2s, 3s, 4s, and 5s.
MIT License
Copyright (c) 2021 Université Paris-Saclay
Copyright (c) 2021 CNRS
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
Logo generated by DesignEvo free logo designer