convert.py

# Converts Amazon's DiPCo dataset to parquet format in an utterance-wise manner
# Author: Benjamin Paine <painebenjamin@gmail.com>
from __future__ import annotations

import os
import gc
import json
import torch
import torchaudio

from time import perf_counter
from datasets import Dataset, Features, Audio, Value, ClassLabel
from typing import Sequence, Dict, Any, Iterable, Iterator, TYPE_CHECKING

if TYPE_CHECKING:
    # Guard this import for type-checking only
    from typing_extensions import Literal

class RotatingAudioDirectoryData:
    """
    A helper class for reading data just-in-time from a directory, and maintaining
    a FIFO-based cache of data in memory.

    Also maintains a loose memory size, evicting last-read data when necessary.
    Accessing data will opaquely read from disk when eneded, with access resetting
    its place in the eviction queue.
    """
    data: Dict[str, torch.Tensor]
    data_sizes: Dict[str, int]
    data_access_times: Dict[str, float]
    sample_rate: int
    eviction_rate: float

    def __init__(self, directory: str, max_size: int, eviction_rate: float=0.25) -> None:
        """
        :param directory: The directory to read data from.
        :param max_size: The maximum size of data to keep in memory.
        :param eviction_rate: The fraction of data to evict when full.
        """
        self.directory = directory
        self.max_size = max_size
        self.file_names = [
            f for f in os.listdir(directory)
            if f.endswith(".wav")
            and not f.startswith(".")
        ]
        self.data = {}
        self.data_sizes = {}
        self.access_times = {}
        self.eviction_rate = eviction_rate
        self.sample_rate = 0
        self.read(self.file_names[0]) # Get sample rate

    @property
    def size(self) -> int:
        """
        :return: The total size of data in memory.
        """
        return sum(self.data_sizes.values())

    def evict(self) -> None:
        """
        Evicts the least-recently accessed items from memory.
        
        This is made to be called infrequently, so it will evict up to the
        configured eviction rate (default 25%) of the total data.
        """
        num_to_evict = int(self.eviction_rate * len(self.data))
        evict_keys = sorted(self.access_times, key=self.access_times.get)[:num_to_evict]
        for key in evict_keys:
            del self.data[key]
            del self.data_sizes[key]
            del self.access_times[key]
        gc.collect()

    def check_evict(self, size: int) -> None:
        """
        Checks if the new data will fit, and evicts if necessary.

        :param size: The size of the new data to add.
        """
        if self.size + size > self.max_size:
            self.evict()

    def read(self, file_name: str) -> None:
        """
        Reads a file from disk and stores it in memory.

        :param file_name: The name of the file to read.
        """
        file_path = os.path.join(self.directory, file_name)
        file_size = os.path.getsize(file_path) * 2 # 16-bit audio
        self.check_evict(file_size)

        try:
            data, sample_rate = torchaudio.load(file_path)
        except RuntimeError as e:
            raise RuntimeError(f"Error reading file {file_path}: {e}")

        if self.sample_rate == 0:
            self.sample_rate = sample_rate

        assert self.sample_rate == sample_rate, "Unexpected sample rate mismatch"

        self.data[file_name] = data[0] # Mono audio
        self.data_sizes[file_name] = file_size
        self.access_times[file_name] = perf_counter()

    def __getitem__(self, key: str) -> torch.Tensor:
        """
        Gets an item from the data, reading it from disk if necessary.

        :param key: The key of the item to get.
        :return: The data corresponding to the key.
        """
        if key not in self.data:
            if key not in self.file_names:
                raise KeyError(f"File {key} not found in directory")
            self.read(key)
        self.access_times[key] = perf_counter()
        return self.data[key]

def maybe_use_tqdm(
    iterable: Iterable[Any],
    desc: Optional[str]=None
) -> Iterator[Any]:
    """
    Uses tqdm if available, otherwise iterates as-is.

    :param iterable: The iterable to iterate over.
    :param desc: The description to show in the progress bar.
    :return: The iterator over the iterable.
    """
    try:
        import tqdm
        yield from tqdm.tqdm(iterable, desc=desc)
    except ImportError:
        yield from iterable

def mix_audio(audio_to_mix: Sequence[torch.Tensor]) -> torch.Tensor:
    """
    Mixes multiple audio arrays together.
    """
    mixed_audio = torch.stack(audio_to_mix)
    return torch.mean(mixed_audio, dim=0)

def get_seconds_from_timestamp(timestamp: str) -> float:
    """
    Converts a timestamp string to seconds.
    Expects a timestamp of format `hh:mm:ss.ff`

    :param timestamp: The timestamp string to convert.
    :return: The number of seconds represented by the timestamp.
    """
    parts = timestamp.split(":")
    hours = int(parts[0])
    minutes = int(parts[1])
    seconds = float(parts[2])
    return hours * 3600 + minutes * 60 + seconds

def process_session_file(
    session_file: str,
    wav_data: RotatingAudioDirectoryData,
    channel_mode: Literal["split", "mixed"] = "split",
) -> Iterable[Dict[str, Any]]:
    """
    Processes a single session file.

    :param session_file: The path to the session file to process.
    :param wav_data: The audio data to use for the session.
    :param channel_mode: The channel mode to use for processing.
    :return: An iterator over the processed utterances.
    """
    with open(session_file, "r") as f:
        session_data = json.load(f)

    # Static device distance data from the DiPCo paper (https://arxiv.org/abs/1909.13447)
    participant_position_device_distance_map = [
        [1600, 2240, 3825, 2900, 1760],
        [1990, 2130, 3950, 3100, 1760],
        [1820, 1520, 2900, 2030, 2790],
        [1300, 1120, 3100, 2520, 2820]
    ]
    # Static map of session to ordered participants
    participant_session_positions = [
        [ 0,  1,  2,  3],
        [ 4,  5,  6,  7],
        [ 8,  9, 10, 11],
        [12, 13, 14, 15],
        [16, 17, 18, 19],
        [20, 21, 22, 23],
        [20, 21, 22, 23],
        [24, 25, 26, 27],
        [28, 29, 30, 31],
        [28, 29, 30, 31]
    ]

    for utterance in maybe_use_tqdm(session_data, desc=f"Processing utterances"):
        # Grab data from dict
        start_times = utterance["start_time"]
        end_times = utterance["end_time"]
        words = utterance["words"]
        gender = utterance["gender"]
        nativeness = utterance["nativeness"]
        mother_tongue = utterance["mother_tongue"]

        session_id_label = utterance["session_id"]
        session_id = int(session_id_label[1:]) - 1

        participant_id_label = utterance["speaker_id"]
        participant_id = int(participant_id_label[1:]) - 1

        # Get the participant's position index and device distances
        participant_position_index = participant_session_positions[session_id].index(participant_id) # 0 - 3
        participant_device_distances = participant_position_device_distance_map[participant_position_index] # 5 distances

        speaker_metadata = {
            "transcription": words,
            "participant_id": participant_id_label,
            "session_id": session_id_label,
            "gender": gender,
            "nativeness": nativeness,
            "mother_tongue": mother_tongue,
        }

        # Each key is either "close-talk" or U01, U02, etc.
        # Go through each and get the timestamp into the corresponding wav file
        for time_key in start_times.keys():
            # Get the start and end times in seconds and frames
            start_timestamp = start_times[time_key]
            start_time_s = get_seconds_from_timestamp(start_timestamp)
            start_frame = int(start_time_s * wav_data.sample_rate)
            # Get the end time in seconds and frames
            end_timestamp = end_times[time_key]
            end_time_s = get_seconds_from_timestamp(end_timestamp)
            end_frame = int(end_time_s * wav_data.sample_rate)
            # Create a metadata dict for the utterance to use later
            utterance_metadata = {
                **speaker_metadata,
                **{
                    "start_timestamp": start_timestamp,
                    "start_time_s": start_time_s,
                    "start_frame": start_frame,
                    "end_timestamp": end_timestamp,
                    "end_time_s": end_time_s,
                    "end_frame": end_frame,
                    "duration_s": end_time_s - start_time_s,
                    "duration_frames": end_frame - start_frame,
                },
            }

            if time_key == "close-talk":
                # Use participant's close-talk audio
                device_metadata = {
                    "device_type": "close-talk",
                    "device_id": participant_id_label,
                    "device_channel": 0,
                    "device_distance_mm": 0,
                }
                wav_file_name = f"{session_id_label}_{participant_id_label}.wav"
                audio_array = wav_data[wav_file_name][start_frame:end_frame]
                yield {
                    "audio": {
                        "array": audio_array,
                        "path": wav_file_name,
                        "sampling_rate": wav_data.sample_rate,
                    },
                    **device_metadata,
                    **utterance_metadata,
                }
            else:
                # Time key will be U01, U02, etc.
                device_id = int(time_key[1:]) - 1
                # Use the far-field audio array of 7 microphones
                device_metadata = {
                    "device_type": "far-field",
                    "device_id": time_key,
                    "device_distance_mm": participant_device_distances[device_id],
                }
                audio_to_mix = []
                # Iterate over each channe
                for channel in range(7):
                    wav_file_name = f"{session_id_label}_{time_key}.CH{channel+1}.wav"
                    audio_array = wav_data[wav_file_name][start_frame:end_frame]

                    if channel_mode == "split":
                        # If using split mode, yield each channel separately
                        yield {
                            "audio": {
                                "array": audio_array,
                                "path": wav_file_name,
                                "sampling_rate": wav_data.sample_rate,
                            },
                            "device_channel": channel+1,
                            **device_metadata,
                            **utterance_metadata,
                        }
                    else:
                        # Otherwise, add to the mix
                        audio_to_mix.append(audio_array)

                if channel_mode == "mixed":
                    # If using mixed mode, mix the audio channels together
                    audio_array = mix_audio(audio_to_mix)
                    yield {
                        "audio": {
                            "array": audio_array,
                            "sampling_rate": wav_data.sample_rate,
                        },
                        "device_channel": 0,
                        **device_metadata,
                        **utterance_metadata,
                    }

def process_split(
    dipco_path: str,
    dipco_split: str,
    channel_mode: Literal["split", "mixed"] = "split",
    max_memory_bytes: int = 16*1024**3, # 16GB
) -> Iterable[Dict[str, Any]]:
    """
    Processes a split of the DiPCo dataset, iterating through all session files.

    :param dipco_path: The path to the DiPCo dataset.
    :param dipco_split: The split of the DiPCo dataset to process.
    :param channel_mode: The channel mode to use for processing.
    :param max_memory_bytes: The maximum memory to use for audio data.
    :return: An iterator over the processed utterances.
    :see: process_session_file
    """
    dipco_path = os.path.abspath(dipco_path)
    wav_dir = os.path.join(dipco_path, "audio", dipco_split)
    wav_data = RotatingAudioDirectoryData(wav_dir, max_size=max_memory_bytes)
    transcriptions_dir = os.path.join(dipco_path, "transcriptions", dipco_split)
    session_filenames = [
        f for f in os.listdir(transcriptions_dir)
        if f.endswith(".json")
        and not f.startswith(".")
    ]

    for session_filename in maybe_use_tqdm(session_filenames, desc=f"Processing session data"):
        num_yielded = 0
        for utterance in process_session_file(
            os.path.join(transcriptions_dir, session_filename),
            wav_data,
            channel_mode=channel_mode,
        ):
            num_yielded += 1
            yield utterance

        print(f"Parsed {num_yielded} utterances from {session_filename}")

    del wav_data
    gc.collect()

def get_split_dataset(
    dipco_path: str,
    dipco_split: str,
    channel_mode: Literal["split", "mixed"] = "split",
) -> Dataset:
    """
    Gets a split of the DiPCo dataset as a Dataset object.

    :param dipco_path: The path to the DiPCo dataset.
    :param dipco_split: The split of the DiPCo dataset to process.
    :param channel_mode: The channel mode to use for processing.
    :return: The processed dataset.
    :see: process_split
    """
    gen_kwargs = {
        "dipco_path": dipco_path,
        "dipco_split": dipco_split,
        "channel_mode": channel_mode,
    }

    return Dataset.from_generator(
        process_split,
        gen_kwargs=gen_kwargs,
        features=Features({
            "audio": Audio(),
            "start_timestamp": Value(dtype="string"),
            "start_time_s": Value(dtype="float32"),
            "start_frame": Value(dtype="uint64"),
            "end_timestamp": Value(dtype="string"),
            "end_time_s": Value(dtype="float32"),
            "end_frame": Value(dtype="uint64"),
            "duration_s": Value(dtype="float32"),
            "duration_frames": Value(dtype="uint64"),
            "transcription": Value(dtype="string"),
            "mother_tongue": Value(dtype="string"),
            "participant_id": Value(dtype="string"),
            "session_id": Value(dtype="string"),
            "device_id": Value(dtype="string"),
            "device_channel": Value(dtype="uint8"),
            "device_distance_mm": Value(dtype="uint16"),
            "device_type": ClassLabel(
                num_classes=2,
                names=["close-talk", "far-field"]
            ),
            "gender": ClassLabel(
                num_classes=2,
                names=["female", "male"]
            ),
            "nativeness": ClassLabel(
                num_classes=2,
                names=["native", "non-native"]
            ),
        })
    )

def synchronize_split(
    dipco_path: str,
    dipco_split: str,
    hub_path: str,
    hub_split: str,
    channel_mode: Literal["split", "mixed"] = "split",
    set_default: bool = False,
) -> None:
    """
    Synchronizes a split of the DiPCo dataset to hub.

    :param dipco_path: The path to the DiPCo dataset.
    :param dipco_split: The split of the DiPCo dataset to process.
    :param hub_path: The path to the hub dataset.
    :param hub_split: The split of the hub dataset to push to.
    :param channel_mode: The channel mode to use for processing.
    :param set_default: Whether to set the split as the default.
    :see: get_split_dataset
    """
    dataset = get_split_dataset(
        dipco_path=dipco_path,
        dipco_split=dipco_split,
        channel_mode=channel_mode,
    )
    dataset.push_to_hub(
        hub_path,
        config_name=f"{channel_mode}-channel",
        split=hub_split,
        set_default=set_default,
    )

# Helper classes and methods done, this is the main process when running the script
if __name__ == "__main__":
    # If you want to run this script, you can configure the following variables
    # to your device and repository settings.
    dipco_path = "./Dipco" # Extracted Dipco dataset
    hub_path = "benjamin-paine/dinner-party-corpus" # username/repo on huggingface.co
    channel_modes = ["split", "mixed"] # channel modes to process, first becomes default
    split_maps = [("dev", "train"), ("eval", "test")] # Map dipco splits to hub splits

    # Done configuring, now run the conversion
    for i, channel_mode in enumerate(channel_modes):
        for j, (dipco_split, hub_split) in enumerate(split_maps):
            synchronize_split(
                dipco_path=dipco_path,
                dipco_split=dipco_split,
                hub_path=hub_path,
                hub_split=hub_split,
                channel_mode=channel_mode,
                set_default=i==0 and j==0,
            )