buckeye_asr.py

# coding=utf-8
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
import logging
from pathlib import Path
import re

import datasets
from datasets.tasks import AutomaticSpeechRecognition


_CITATION = """\
@misc{pitt2007Buckeye,
title = {Buckeye {Corpus} of {Conversational} {Speech} (2nd release).},
url = {www.buckeyecorpus.osu.edu},
publisher = {Columbus, OH: Department of Psychology, Ohio State University (Distributor)},
author = {Pitt, M.A. and Dilley, L. and Johnson, K. and Kiesling, S. and Raymond, W. and Hume, E. and Fosler-Lussier, E.},
year = {2007},
}
"""

_DESCRIPTION = """\
The Buckeye Corpus of conversational speech contains high-quality recordings
from 40 speakers in Columbus OH conversing freely with an interviewer. The
speech has been orthographically transcribed and phonetically labeled.
"""
_HOMEPAGE = "https://buckeyecorpus.osu.edu"

_LICENSE = "FREE for noncommercial uses"


class BuckeyeASRDataset(datasets.GeneratorBasedBuilder):
    """Buckeye ASR dataset."""

    VERSION = datasets.Version("1.0.0")
    PHONE_EXCLUSION_LIST = ["{B_TRANS}", "<EXCLUDE-name>", "IVER", "NOISE", "SIL", "VOCNOISE"]
    SPLIT_TAGS = ['SIL', 'IVER']

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "file": datasets.Value("string"),
                    "audio": datasets.Value("string"),
                    "text": datasets.Value("string"),
                    "phonetic_detail": datasets.Sequence(
                        {
                            "start": datasets.Value("float"),
                            "stop": datasets.Value("float"),
                            "label": datasets.Value("string"),
                        }
                    ),
                    "word_detail": datasets.Sequence(
                        {
                            "start": datasets.Value("float"),
                            "stop": datasets.Value("float"),
                            "label": datasets.Value("string"),
                            "broad_transcription": datasets.Value("string"),
                            "narrow_transcription": datasets.Value("string"),
                            "syntactic_class": datasets.Value("string"),
                        }
                    ),
                    "start": datasets.Value("float"),
                    "stop": datasets.Value("float"),
                    "speaker_id": datasets.Value("string"),
                    "id": datasets.Value("string"),
                }
            ),
            supervised_keys=("file", "text"),
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
            task_templates=[AutomaticSpeechRecognition(
                audio_file_path_column="file", transcription_column="text")],
        )

    def _split_generators(self, dl_manager):
        """Returns SplitGenerators."""
        # We use 10% of the data for validation and test sets, with no speaker
        # overlap between the sets. We also balance for speaker's gender and
        # age.
        spkr_ids_dev = [18, 34, 36, 37]
        spkr_ids_test = [20, 38, 39, 40]

        def is_training(i):
            return i not in spkr_ids_dev and i not in spkr_ids_test
        spkr_ids_train = filter(is_training, range(1, 41))

        paths = {
            "train": [f"{dl_manager.manual_dir}/s{i:02n}" for i in spkr_ids_train],
            "dev": [f"{dl_manager.manual_dir}/s{i:02n}" for i in spkr_ids_dev],
            "test": [f"{dl_manager.manual_dir}/s{i:02n}" for i in spkr_ids_test],
        }
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"paths": paths["train"]},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"paths": paths["test"]},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={"paths": paths["dev"]},
            ),
        ]

    def _generate_examples(self, paths):
        """ Yields examples as (key, example) tuples. """
        for p in paths:
            for wav_path in Path(p).glob("*.wav"):
                # Extract words
                fpath = wav_path.with_suffix(".words")
                wordlist = self._extract_word_info(fpath)
                word_seqs = self._split_words(wordlist, fpath)

                # Extract transcript
                # To avoid conflict between the transcripts (`.txt` files) and
                # the word alignment files (`.words` files), we compute the
                # transcripts from the latter.
                transcripts = [" ".join([w["label"] for w in wseq]) for wseq in word_seqs]

                # Extract phones
                fpath = wav_path.with_suffix(".phones")
                phonelist = self._extract_phone_info(fpath)
                phone_seqs = self._split_phones(word_seqs, phonelist, fpath)
                assert len(phone_seqs) == len(word_seqs)

                # id_ must be a unique key
                for idx in range(len(transcripts)):
                    ws = word_seqs[idx]
                    id_ = f"{wav_path.stem}_{idx}"
                    if phone_seqs[idx]:
                        start = min(ws[0]["start"], phone_seqs[idx][0]["start"])
                        stop = max(ws[-1]["stop"], phone_seqs[idx][-1]["stop"])
                    else:
                        start = ws[0]["start"]
                        stop = ws[-1]["stop"]
                    shift_timestamps(ws, -start)
                    shift_timestamps(phone_seqs[idx], -start)
                    example = {
                        "file": wav_path,
                        "audio": wav_path,
                        "text": transcripts[idx],
                        "phonetic_detail": phone_seqs[idx],
                        "word_detail": ws,
                        "speaker_id": id_[1:3],
                        "id": id_,
                        "start": start,
                        "stop": stop,
                    }
                    yield id_, example

    def _extract_word_info(cls, fpath):
        with open(fpath) as f:
            lines = f.readlines()
        start = 0
        wordlist = []
        for line in lines[9:]:
            # File s1901b.words goes beyond the end of the audio
            if fpath.name == "s1901b.words" and start > 568.739:
                break
            line = line.rstrip("\n")
            if not line:  # Skipping empty lines
                continue
            fields = line.split("; ")
            subfields = fields[0].split()
            label = subfields[2]
            stop = float(subfields[0])
            if label[0] in ['<', '{']:
                # Handling tags (tags sometime miss transcriptions)
                wordlist.append({
                    "start": start,
                    "stop": stop,
                    "label": label,
                })
            else:
                # Handling words
                if len(fields) < 4:
                    logging.warning(f"Line \"{line}\" missing fields in file {fpath}")
                else:
                    narrow_trn = fields[2]
                    # Warning if the narrow_transcription is empty
                    if not narrow_trn:
                        logging.warning(f"Narrow transcription is empty in {fpath}")
                    wordlist.append({
                        "start": start,
                        "stop": stop,
                        "label": label,
                        "broad_transcription": fields[1],
                        "narrow_transcription": narrow_trn,
                        "syntactic_class": fields[3],
                    })
            start = stop
        return wordlist

    def _split_words(cls, wordlist, fpath):
        word_seqs = []
        segment = []
        for w in wordlist:
            if w["label"][0] in ['<', '{']:
                # Segmenting if the label starts with one of the SPLIT_TAGS
                regexp = "|".join([f"<{st}" for st in cls.SPLIT_TAGS])
                match = re.match(regexp, w["label"])
                if match and match.start() == 0 and segment:
                    # The model can't handle segments shorter than 0.25 ms
                    if segment[-1]["stop"] - segment[0]["start"] >= 0.025:
                        word_seqs.append(segment)
                    else:
                        logging.warning(
                            f"Sequence shorter than 25 ms in {fpath} starting "
                            f"at {segment[0]['start']}")
                    segment = []
            else:
                segment.append(w)
        if segment:
            if segment[-1]["stop"] - segment[0]["start"] >= 0.025:
                word_seqs.append(segment)
            else:
                logging.warning(
                    f"Sequence shorter than 25 ms in {fpath} starting "
                    f"at {segment[0]['start']}")
        return word_seqs

    def _extract_phone_info(cls, fpath):
        with open(fpath) as f:
            lines = f.readlines()
        start = 0
        phonelist = []
        for i, line in enumerate(lines[9:]):
            if line == '\n':
                # empty line -> skip
                continue
            # some lines contain "; *" after the last field
            # we should get rid of that
            line = line.split(";")[0].strip()
            fields = line.split()
            stop = float(fields[0])
            if len(fields) < 3:
                # phone without label
                logging.warning(
                    f"Phone missing label in file {fpath} at time {stop}.")
                phone = None
            else:
                phone = fields[2]
            phonelist.append({
                "start": start,
                "stop": stop,
                "label": phone,
            })
            start = stop
        return phonelist

    # Matching phone and word alignments based on exact timestamps doesn't
    # work as the timestamps don't always match.
    # Also, the narrow transcription can conflict with the phone alignments
    # in many ways, including:
    #  * wrong phone label in one of the two (e.g. `tq` instead of `t`).
    #  * phone absent from the narrow transcription can appear in the
    #    alignments with no duration.
    #  * if two consecutive phones are the same in the narrow transcription,
    #    they can be merged in the alignements.
    # We are conservative here and keep every phone that is completely included
    # in the utterance + overlapping starting/ending phones if they match the
    # first/last phone of the narrow transcription.
    def _split_phones(cls, word_seqs, phonelist, fpath):
        phone_seqs = []
        i_pl = 0
        for wl in word_seqs:
            segment = []
            # skip phones that precede the first word's start
            while phonelist[i_pl]["stop"] <= wl[0]["start"]:
                i_pl += 1
            # a phone which overlaps with the start of the first word might
            # not belong with the utterance
            if (phonelist[i_pl]["label"] and
                    phonelist[i_pl]["start"] < wl[0]["start"] and
                    not wl[0]["narrow_transcription"].startswith(phonelist[i_pl]["label"])):
                # skip the phone
                i_pl += 1
            # take all phones fully included in the utterance
            while (i_pl < len(phonelist) and
                   phonelist[i_pl]["stop"] < wl[-1]["stop"]):
                segment.append(phonelist[i_pl])
                i_pl += 1
            # a phone which overlaps with the end of the last word might
            # belong with the utterance
            if (phonelist[i_pl]["label"] and
                    phonelist[i_pl]["start"] < wl[-1]["stop"] and
                    wl[-1]["narrow_transcription"].endswith(phonelist[i_pl]["label"])):
                # add the phone
                segment.append(phonelist[i_pl])
                i_pl += 1
            phone_seqs.append(segment)
        return phone_seqs


def shift_timestamps(detail, shift):
    for item in detail:
        item["start"] += shift
        item["stop"] += shift

def matching(phone, word):
    # Reject phones without label
    if phone["label"] is None:
        return False
    # Phones that are totally encompassed by the word timewise are considered
    # part of the word
    if phone["start"] >= word["start"] and phone["stop"] <= word["stop"]:
        return True
    # In case of overlap with the start or the end, we further request a time
    # difference <= 20 ms and check the label
    if phone["start"] < word["start"]:
        # Possible overlap with the start
        if (included(phone, word) and
                word["narrow_transcription"].startswith(phone["label"])):
            return True
    # Last option is an overlap with the end of the word
    return (included(phone, word) and
            word["narrow_transcription"].endswith(phone["label"]))


def included(phone, word, threshold=0.02):
    # We accept an overlap with time difference up to the threshold at the
    # start or the end
    return (phone["start"] >= word["start"] - threshold and
            phone["start"] < word["stop"] and
            phone["stop"] > word["start"] and
            phone["stop"] <= word["stop"] + threshold)