- Create environment:
conda create -n diart python=3.8
conda activate diart- Install
PortAudioandsoundfile:
conda install portaudio
conda install pysoundfile -c conda-forge- Install diart:
pip install diartBy default, diart is based on pyannote.audio models stored in the huggingface hub. To allow diart to use them, you need to follow these steps:
- Accept user conditions for the
pyannote/segmentationmodel - Accept user conditions for the
pyannote/embeddingmodel - Install huggingface-cli and log in with your user access token (or provide it manually in diart CLI or API).
A recorded conversation:
diart.stream /path/to/audio.wavA live conversation:
diart.stream microphoneSee 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 import OnlineSpeakerDiarization
from diart.sources import MicrophoneAudioSource
from diart.inference import RealTimeInference
from diart.sinks import RTTMWriter
pipeline = OnlineSpeakerDiarization()
mic = MicrophoneAudioSource(pipeline.config.sample_rate)
inference = RealTimeInference(pipeline, mic, do_plot=True)
inference.attach_observers(RTTMWriter(mic.uri, "/output/file.rttm"))
prediction = 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 import OnlineSpeakerDiarization, PipelineConfig
from diart.models import EmbeddingModel
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)
prediction = 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/dirSee 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/exampleor 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.rearrange_audio_stream(sample_rate=sample_rate),
ops.map(lambda wav: (wav, segmentation(wav))),
ops.starmap(embedding)
).subscribe(on_next=lambda emb: print(emb.shape))
mic.read()Output:
# Shape is (batch_size, num_speakers, embedding_dim)
torch.Size([1, 3, 512])
torch.Size([1, 3, 512])
torch.Size([1, 3, 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 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()))
prediction = 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@Interspeech2021or using the inference API:
from diart.inference import Benchmark
from diart 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
-
Target & Code function this project is about recognizing the voice in a meeting, conversations, and.... whenever someone talks, the detectors can find and detect which voice belongs to which person. however it can find the muted part in conversation and omitted. finding the voice and recognizing can happened withoud overlapping. even more than 2 person talk in same time. we use GMM cluster and SAD to find the voice source. the input ofcourse is voice and the outputs are numbers and results in papers.
-
describe innovation
you can find the inovation files among the other files and I describe it even in my video.
- changes in codes
Speakers diarization with detecting the latency and overlap of online speakers The project has been based on MIT databases for training network. In order to compile the desired code, the following stages must be taken: 1.install pycharm 2.install pip python --upgrade the python version =3.9 (it is necessary) 3. pip install pip --upgrade 4. install anaconda (it takes time) after the anaconda is installed the environment in conda based on python 3.9 is made. 5.the below libraries should be add in order to use them in code. diart, tourch, tourchaudio, pyannote.audio 6.conda create -n diart python=3.9 7.conda activate diart 8.conda create -n diart python=3.8 9.conda activate diart 10.pip install diart 11.pyannote.audio should be installed 12. the terms of MIT have to be consented by user by following the https://huggingface.co/pyannote/segmentation 13. order the token for login open the below link: 14. pip install huggingface_hub 15.conda install -c conda-forge huggingface_hub 16. use the token code for logging to the MIT database 17.the main.py must be run 18. please enter the token number 19. after compile the main.py 20. the script turn the laptop's speaker on and the voice of speaker are saved every 0.5 in buffers. 21. by using pipeline all voices are recorded and according to the frequency the speaker voices is recognized. 22. the network will be trained by 5 sec recording 23. the crucial parts of this project are latency and the over-lap of speakers' voices. the problems are solved by separating the frequency according to the trained network and pytourch library.
4.results
you can see the results in the files amoung the other files uploaded. also, the result contains:
In oreder to compile the desired code the following stages must be taken: 1.install pycharm 2.install pip python --upgrade the python version =3.8 ( it is necessary) 3. pip instal pip --upgrade 4. inistall anaconda ( it takes time) after the anaconda is inistalled the environment in conda based on python 3.8 is made. 5.the below libraries should be add in order to use them in code. diart,tourch,tourchaudio,py 6.conda create -n diart python=3.9 7.conda activate diart 8.conda create -n diart python=3.8 9.conda activate diart 10.pip install diart 11.pyannote.audio should be installed 12. the terms of MIT have to be consented by user by following the https://huggingface.co/pyannote/segmentation 13. order the token for login open the below link: 14. pip install huggingface_hub 15.conda install -c conda-forge huggingface_hub 16. use the token code for logging to the MIT database 17.the main.py must be run 18. please enter the token number 19. after compile the main.py 20. the script turn the laptop's speaker on and the voice of speaker are saved every 0.5 in buffers. 21. by using pipeline all voices are recorded and according to the frequency the speaker voices is regognize. 22. the network will be trained by 5 sec recording 23. the crucial part of this project is
- Reference
you can find the reference paper amoung the files uploaded.
- Interduce myself
this is Narges Pourakhlaghi, Master student of Biomedical engineering- Bioelectric, in Islamic Azad University South Tehran Branch. this project done for DSP-Dr. Mahdi Eslami, first semester of 1401
7.Paper file uploaded
8.explaining video
https://drive.google.com/file/d/1WM5v1-DJ5ZNK8mc_X8YUzLWRRy-BhOkp/view?usp=share_link
9.Proposal file uploaded