"""
Utilities for file format conversion for Speaker Diarization
Authors:
* Jiatong Shi 2021
* Leo 2022
"""
import logging
import os
import re
import subprocess
from pathlib import Path
from typing import List
import numpy as np
import pandas as pd
import torch
from scipy.signal import medfilt
from tqdm import tqdm
logger = logging.getLogger(__name__)
RTTM_FORMAT = "SPEAKER {:s} 1 {:7.2f} {:7.2f} <NA> <NA> {:s} <NA>"
__all__ = [
"kaldi_dir_to_rttm",
"csv_to_kaldi_dir",
"kaldi_dir_to_csv",
]
[docs]def kaldi_dir_to_rttm(data_dir: str, rttm_path: str):
data_dir: Path = Path(data_dir)
segments_file = data_dir / "segments"
utt2spk_file = data_dir / "utt2spk"
assert segments_file.is_file()
assert utt2spk_file.is_file()
utt2spk = {}
with utt2spk_file.open() as f:
for utt2spk_line in f.readlines():
fields = utt2spk_line.strip().replace("\n", " ").split()
assert len(fields) == 2
utt, spk = fields
utt2spk[utt] = spk
with Path(rttm_path).open("w") as rttm_f:
with segments_file.open() as f:
for segment_line in f.readlines():
fields = segment_line.strip().replace("\t", " ").split()
assert len(fields) == 4
utt, reco, start, end = fields
spk = utt2spk[utt]
print(
RTTM_FORMAT.format(
reco,
float(start),
float(end) - float(start),
spk,
),
file=rttm_f,
)
def make_rttm_and_score(
prediction_dir: str,
score_dir: str,
gt_rttm: str,
frame_shift: int,
thresholds: List[int],
medians: List[int],
subsampling: int = 1,
sampling_rate: int = 16000,
):
Path(score_dir).mkdir(exist_ok=True, parents=True)
dscore_dir = Path(score_dir) / "dscore"
rttm_dir = Path(score_dir) / "rttm"
result_dir = Path(score_dir) / "result"
setting2dscore = []
for th in thresholds:
for med in medians:
logger.info(f"Make RTTM with threshold {th}, median filter {med}")
rttm_file = rttm_dir / f"threshold-{th}_median-{med}.rttm"
make_rttm(
prediction_dir,
rttm_file,
th,
med,
frame_shift,
subsampling,
sampling_rate,
)
logger.info(f"Scoring...")
result_file = result_dir / f"threshold-{th}_median-{med}.result"
overall_der = score_with_dscore(dscore_dir, rttm_file, gt_rttm, result_file)
logger.info(f"DER: {overall_der}")
setting2dscore.append(((th, med), overall_der))
setting2dscore.sort(key=lambda x: x[1])
(best_th, best_med), best_der = setting2dscore[0]
return best_der, (best_th, best_med)
def make_rttm(
prediction_dir: str,
out_rttm_path: str,
threshold: int,
median: int,
frame_shift: int,
subsampling: int,
sampling_rate: int,
):
names = sorted([name for name in os.listdir(prediction_dir)])
filepaths = [Path(prediction_dir) / name for name in names]
Path(out_rttm_path).parent.mkdir(exist_ok=True, parents=True)
with open(out_rttm_path, "w") as wf:
for filepath in filepaths:
session, _ = os.path.splitext(os.path.basename(filepath))
data = torch.load(filepath).numpy()
a = np.where(data > threshold, 1, 0)
if median > 1:
a = medfilt(a, (median, 1))
factor = frame_shift * subsampling / sampling_rate
for spkid, frames in enumerate(a.T):
frames = np.pad(frames, (1, 1), "constant")
(changes,) = np.where(np.diff(frames, axis=0) != 0)
for s, e in zip(changes[::2], changes[1::2]):
print(
RTTM_FORMAT.format(
session,
s * factor,
(e - s) * factor,
session + "_" + str(spkid),
),
file=wf,
)
def score_with_dscore(
dscore_dir: str, hyp_rttm: str, gt_rttm: str, score_result: str
) -> float:
"""
This function returns the overall DER score, and will also write the detailed scoring results
to 'score_result'
"""
dscore_dir: Path = Path(dscore_dir)
Path(score_result).parent.mkdir(exist_ok=True, parents=True)
if not dscore_dir.is_dir():
logger.info(f"Cloning dscore into {dscore_dir}")
subprocess.check_output(
f"git clone https://github.com/nryant/dscore.git {dscore_dir}",
shell=True,
).decode("utf-8")
subprocess.check_call(
f"python3 {dscore_dir}/score.py -r {gt_rttm} -s {hyp_rttm} > {score_result}",
shell=True,
)
return get_overall_der_from_dscore_file(score_result)
def get_overall_der_from_dscore_file(score_result: str):
with open(score_result) as file:
lines = file.readlines()
overall_lines = [line for line in lines if "OVERALL" in line]
assert len(overall_lines) == 1
overall_line = overall_lines[0]
overall_line = re.sub("\t+", " ", overall_line)
overall_line = re.sub(" +", " ", overall_line)
overall_der = float(overall_line.split(" ")[3])
# The overall der line should look like:
# *** OVERALL *** DER JER ...
return overall_der
[docs]def csv_to_kaldi_dir(csv: str, data_dir: str):
logger.info(f"Convert csv {csv} into kaldi data directory {data_dir}")
data_dir: Path = Path(data_dir)
data_dir.mkdir(exist_ok=True, parents=True)
df = pd.read_csv(csv)
required = ["record_id", "wav_path", "utt_id", "speaker", "start_sec", "end_sec"]
for r in required:
assert r in df.columns
reco2path = {}
reco2dur = {}
utt2spk = {}
spk2utt = {}
segments = []
for rowid, row in tqdm(df.iterrows(), total=len(df)):
record_id, wav_path, duration, utt_id, speaker, start_sec, end_sec = (
row["record_id"],
row["wav_path"],
row["duration"],
row["utt_id"],
row["speaker"],
row["start_sec"],
row["end_sec"],
)
if record_id in reco2path:
assert wav_path == reco2path[record_id]
else:
reco2path[record_id] = wav_path
if record_id not in reco2dur:
reco2dur[record_id] = duration
else:
assert reco2dur[record_id] == duration
if utt_id not in utt2spk:
utt2spk[utt_id] = str(speaker)
else:
assert utt2spk[utt_id] == str(speaker)
if speaker not in spk2utt:
spk2utt[speaker] = []
spk2utt[speaker].append(utt_id)
segments.append((utt_id, record_id, str(start_sec), str(end_sec)))
with (data_dir / "wav.scp").open("w") as f:
f.writelines([f"{reco} {path}\n" for reco, path in reco2path.items()])
with (data_dir / "reco2dur").open("w") as f:
f.writelines([f"{reco} {dur}\n" for reco, dur in reco2dur.items()])
with (data_dir / "utt2spk").open("w") as f:
f.writelines([f"{utt} {spk}\n" for utt, spk in utt2spk.items()])
with (data_dir / "spk2utt").open("w") as f:
f.writelines([f"{spk} {' '.join(utts)}\n" for spk, utts in spk2utt.items()])
with (data_dir / "segments").open("w") as f:
f.writelines(
[f"{utt} {record} {start} {end}\n" for utt, record, start, end in segments]
)
[docs]def kaldi_dir_to_csv(data_dir: str, csv: str):
logger.info(f"Convert kaldi data directory {data_dir} into csv {csv}")
data_dir: Path = Path(data_dir)
assert (data_dir / "wav.scp").is_file()
assert (data_dir / "segments").is_file()
assert (data_dir / "utt2spk").is_file()
assert (data_dir / "reco2dur").is_file()
reco2path = {}
with (data_dir / "wav.scp").open() as f:
for line in f.readlines():
line = line.strip()
reco, path = line.split(" ")
reco2path[reco] = path
reco2dur = {}
with (data_dir / "reco2dur").open() as f:
for line in f.readlines():
line = line.strip()
reco, duration = line.split(" ")
reco2dur[reco] = float(duration)
utt2spk = {}
with (data_dir / "utt2spk").open() as f:
for line in f.readlines():
line = line.strip()
utt, spk = line.split(" ")
utt2spk[utt] = spk
row = []
with (data_dir / "segments").open("r") as f:
for line in f.readlines():
line = line.strip()
utt, reco, start, end = line.split(" ")
row.append(
(
reco,
reco2path[reco],
reco2dur[reco],
utt,
utt2spk[utt],
float(start),
float(end),
)
)
recos, wav_paths, durations, utts, spks, starts, ends = zip(*row)
pd.DataFrame(
data=dict(
record_id=recos,
wav_path=wav_paths,
utt_id=utts,
speaker=spks,
start_sec=starts,
end_sec=ends,
duration=durations,
)
).to_csv(csv, index=False)