Skip to content

Latest commit

 

History

History
266 lines (199 loc) · 8.26 KB

README.md

File metadata and controls

266 lines (199 loc) · 8.26 KB
Raw

ASVspoof2021 Evaluation Package

By ASVspoof2021 challenge organizers

With the release of the full set of keys and meta-labels (see ASVspoof.org), we provide this updated evaluation package to compute min t-DCF and EER.

Compared with the previous evaluation package (archived-package-stage-1), this evaluation package

  • downloads and uses the full set of keys and meta-labels,
  • computes not only pooled but also decomposed min t-DCFs and EERs on specified conditions,
  • allows the users to provide their own ASV scores.

Users are encouraged to use this evaluation package rather than package-stage-1.

Link to keys and meta-label files

Link MD5
LA https://www.asvspoof.org/asvspoof2021/LA-keys-full.tar.gz 037592a0515971bbd0fa3bff2bad4abc
PA https://www.asvspoof.org/asvspoof2021/PA-keys-full.tar.gz a639ea472cf4fb564a62fbc7383c24cf
DF https://www.asvspoof.org/asvspoof2021/DF-keys-full.tar.gz dabbc5628de4fcef53036c99ac7ab93a

You can manually download them.

On Linux system, you may also use download.sh to download them.

Key & Meta-label file is in text format, each line contains the key and meta-labels for one trial. Details of the meta-labels are explained in the ASVspoof 2021 long summary paper.

We also provide a short explanation at the end of this page.

How to use evaluation scripts

You can use the Python scripts to compute EERs and min t-DCFs.

Step 1. Install requirement

pip install numpy
pip install pandas
pip install matplotlib

Step 2. Download keys and meta-labels

Either use

bash download.sh

or manually download and untar them.

A directory called ./keys will be available. It contains:

   keys
   |- LA                            # Files for LA track
   |  |- CM 
   |  |   |- trial_metadata.txt     # CM protocol with keys and meta-labels
   |  |   |- LFCC-GMM
   |  |   |    |- score.txt         # Score file from a baseline LFCC-GMM 
   |  |   |- ...
   |  |
   |  |- ASV
   |      |- trial_metadata.txt     # ASV protocol with keys and meta-labels
   |      |- ASVtorch_kaldi
   |          |- score.txt          # Score file from the ASV system
   |- DF ...
   |- PA ...

Step 3. Compute EER and min t-DCF

A help message can be found by

python main.py --help

Here are some example use cases. Let's assume we have a CM score file score.txt for LA, and we want to get the results on eval subset (i.e., evaluation subset, which is disjoint from the progress and hidden subsets).

Case 1 (most common use case)

Compute results using pre-computed t-DCF C012 coefficients provided by the organizers

python main.py --cm-score-file score.txt --track LA --subset eval

Case 2

Recompute C012 using official ASV scores, save it to ./LA-c012.npy, and use the C012 coefficients to compute min tDCFs

python main.py --cm-score-file score.txt --track LA --subset eval --recompute-c012 --c012-path ./LA-c012.npy

Case 3

Recompute C012 using my own ASV scores, save it to ./LA-c012.npy and use the new C012 to compute min tDCFs

python main.py --cm-score-file score.txt --track LA --subset eval --recompute-c012 --c012-path ./LA-c012.npy --asv-score-file ./asv-score.txt

Case 4

Compute min tDCF using my own pre-computed C012 coeffs ./LA-c012.npy

python main.py --cm-score-file score.txt --track LA --subset eval --c012-path ./LA-c012.npy

If you don't have score.txt at hand

You may play with the code using baseline CM score files.

They are available in the downloaded key and meta-label file packages

ls keys/*/CM/*/score.txt

How to use notebook

Based on the Python scripts, this interactive notebook shows the details of min t-DCF and EER computation.

It also includes an API, which allows the user to upload score file and get the min t-DCF and EER tables.

You can directly open it through Google Colab. Just click the badge Open In Colab

On meta-labels

Here we briefly explain the meanings of meta-labels, using the first line in LA/CM/trial_metadata.txt, PF/CM/trial_metadata.txt, and DF/CM/trial_metadata.txt.

LA

LA_0009 LA_E_9332881 alaw ita_tx A07 spoof notrim eval
  • LA_0009: speaker ID
  • LA_E_9332881: trial ID
  • alaw: name of codec. It can be:
    • none: LA-C1
    • alaw: LA-C2
    • pstn: LA-C3
    • g722: LA-C4
    • ulaw: LA-C5
    • gsm : LA-C6
    • opus: LA-C7
  • ita_tx: name of transmission condition. It can be
    • ita_tx: FR-IT, transmission between France and Italy
    • sin_tx: FR-SG, transmission between France and Singapore
    • loc_tx: local transmission
    • mad_tx: Transmission through PSTN to Spain
  • A07: name of spoofing attack. It can be
    • A07 - A19 are defined in ASVspoof 2019 LA database
  • spoof: key. It can be
    • bonafide: bona fide
    • spoof: spoof
  • notrim: whether non-speech frames are trimmed. It can be
    • notrim: not trimmed
    • trim: trimmed
  • eval: name of subset. It can be
    • eval: evaluation subset
    • progress: progress subset
    • hidden: hidden subset (there non-speech frames are trimmed)

PA

PA_0010 PA_E_1000001 R3 M3 d4 r1 m1 s4 c4 spoof notrim eval
  • PA_0010: speaker ID
  • PA_E_1000001: trial ID
  • Environment factors:
    • R1 - R9: ASV Room IDs
    • M1 - M3: ASV microphone IDs
    • D1 - D6: Talker-to-ASV Distance distances
  • Attacker factors:
    • r1 - r9: Attacker Room IDs
    • m1 - m3: Attacker microphone IDs
    • c2 - c4: Attacker to talker distances
    • s2 - s4: Attacker replay device IDs
    • d1 - d6: Attacker-replay-device-to-ASV distances
  • spoof: key. It can be
    • bonafide: bona fide
    • spoof: spoof
  • notrim: whether non-speech frames are trimmed. It can be
    • notrim: not trimmed
    • trim: trimmed
  • eval: name of subset. It can be
    • eval: evaluation subset
    • progress: progress subset
    • hidden: hidden subsets

Note that hidden subsets contain:

  • notrim hidden: hidden subset 1 that contains simulated trials without trimming
  • trim hidden: hidden subset 2 that contains real-replayed but trimmed trials

Note that, compared with key file released previously, these PA meta-labels are slightly updated:

  • Old notation:
  d2 - d4: Attacker to talker distances
  D1 - D6: Attacker-replay-device-to-ASV distances
  • New notation in the full set of key meta-labels
  c2 - c4: Attacker to talker distances
  d1 - d6: Attacker-replay-device-to-ASV distances

DF

LA_0023 DF_E_2000011 nocodec asvspoof A14 spoof notrim progress traditional_vocoder - - - -
  • LA_0009: speaker ID
  • DF_E_2000011: trial ID
  • nocodec: name of codec for compression. It can be:
    • nocodec: DF-C1
    • low_mp3: DF-C2
    • high_mp3: DF-C3
    • low_m4a: DF-C4
    • high_m4a: DF-C5
    • low_ogg : DF-C6
    • high_ogg: DF-C7
    • mp3m4a : DF-C8
    • oggm4a: DF-C9
  • asvspoof: source of data. It can be:
    • asvspoof: from ASVspoof 2019
    • vcc2018: from VCC 2018
    • vcc2020: from VCC 2020
  • A14: name of spoofing attack
    • A07 - A19 are defined in ASVspoof 2019 LA database
  • spoof: key
    • bonafide: bona fide
    • spoof: spoof
  • notrim: whether non-speech frames are trimmed
    • notrim: not trimmed
    • trim: trimmed
  • progress: name of subset, which can be
    • eval: evaluation subset
    • progress: progress subset
    • hidden: hidden subset (there non-speech frames are trimmed)
  • traditional_vocoder: type of vocoder
    • bonafide: this is a bona fide trial
    • neural_vocoder_autoregressive: spoofed trial using neural AR vocoder
    • neural_vocoder_nonautoregressive: spoofed trial using neural non-AR vocoder
    • traditional_vocoder: spoofed trial using traditional DSP-based vocoder
    • unknown: spoofed trial with an unknown/unannotated vocoder
    • waveform_concatenation: spoofed trial by waveform concatenation

End