Datasets:

sammlapp
/

pa_forest_bird_localization_2023

"""
use the PPS data generated by AudioMothGPSDeploy_1_0_8_Hardware_1_1
to generate audio files starting at
a known time and having the exact desired sampling rate
"""

import pandas as pd
import numpy as np
from scipy.interpolate import interp1d
from datetime import datetime, timedelta
import pytz
from opensoundscape import Audio
import gc
from aru_metadata_parser.parse import parse_audiomoth_metadata_from_path


def parse_RMC_time_to_seconds(timestamp: str) -> float:
    """
    Takes in the GPS-audiomoth formatted timestamp and returns unix epoch time in seconds.
    Args:
        timestamp: string in the format used by our GPS-audiomoths
                '%Y-%m-%dT%H:%M:%S.%f' e.g. "2022-05-14T10:30:00.725"
    Returns: time from unix-epoch in seconds (to 10^-6s resolution)as a float
    """
    stamp_format = "%Y-%m-%dT%H:%M:%S.%f"
    return datetime.strptime(timestamp, stamp_format).timestamp()


def parse_RMC_time_to_datetime(timestamp: str) -> datetime:
    """
    Takes in the GPS-audiomoth formatted timestamp and returns datetime.datetime
    localized with UTC timezone. Assumes the input time string is given in UTC.
    Args:
        timestamp: string in the format used by our GPS-audiomoths
                '%Y-%m-%dT%H:%M:%S.%f' e.g. "2022-05-14T10:30:00.725"
    Returns: datetime object localized to UTC timezone
    """
    stamp_format = "%Y-%m-%dT%H:%M:%S.%f"
    tz = pytz.timezone("UTC")
    return tz.localize(datetime.strptime(timestamp, stamp_format))


def process_pps_file(pps_file, expected_sr=48000, sr_tol=50) -> pd.DataFrame:
    """
    Using the PPS File, creates a table aligning GPS_TIME_SEC values with the corresponding PPS and SAMPLES_WRITTEN Values
    This is the information we need to interpolate audio files to achieve "synchronized" files with the exact desired sampling rate
    This function performs several checks to try to determine whether a value in the LAST_RMC_GPS_TIME column corresponds to the
    PPS from the previous line. If so, the previous line's SAMPLES_WRITTEN value gives the number of samples written at the time of
    the PPS, whose accurate time stamp is given by the LAST_RMC_GPS_TIME of the subsequent row.
    If any of the validation checks fail, we discard the RMC data from that row. As long as there is eventually a good pairing of
    RMC data (LAST_RMC_GPS_TIME) with a PPS, we will know the number of samples written and the exact time interval. This allows us
    to interpolate to achieve the nominal sampling rate.
    Args:
        - pps_file: path to the pps file (comma separated file produced by the firmware `AudioMothGPSDeploy_1_0_8_Hardware_1_1`)
    Returns:
        - a DataFrame with rows in which GPS_TIME_SEC is matched with the corresponding PPS, SAMPLES_WRITTEN, and COUNTER values
        columns: ["PPS","SAMPLES_WRITTEN", "GPS_TIME_SEC", "GPS_TIME_STR", "CPU_CLOCK_COUNTER"]
    """
    pps_table = pd.read_csv(pps_file, index_col=0)

    samples_written = []
    gps_times_str = []
    gps_times_sec = []
    pps_numbers_out = []
    counter_positions = []

    pps_pings_with_no_RMC = []

    # Buffer overflow check: Fail if any buffer overflows occurred
    # a buffer overflow can be detected by a permanent difference of a multiple of 8 between
    # BUFFERS_FILLED and BUFFERS_WRITTEN. A difference of 8 corresponds to 2^17 samples that were
    # recorded from the sensor but never written to the WAV file. Thus TOTAL_SAMPLES is incorrect,
    # it overstates the number of samples actually written to the WAV file (it accurately reports the
    # number of samples recorded from the sensor into the buffer).
    # It may be possible to correct for buffer overflows by inserting 0s at the appropriate positions
    # in the WAV file, but here we fail rather than trying to correct buffer overflows.
    final_row = pps_table.iloc[-1]
    assert (
        final_row["BUFFERS_FILLED"] - final_row["BUFFERS_WRITTEN"] < 8
    ), "Buffer overflow detected, WAV file has missing content"

    pps_numbers = pps_table.index

    # PPS Duplicate Check: check for duplicates in PPS number
    if sum(pps_numbers.duplicated()) > 0:
        raise ValueError(
            f"PPS numbers are duplicated in this file. Something may be wrong: {pps_file}"
        )

    # PPS Sequence Check: check that PPS numbers are sequential from first to last
    assert np.array_equal(
        pps_numbers.values, list(range(pps_numbers[0], pps_numbers[-1] + 1))
    ), "PPS Numbers were not sequential integers"

    for i, pps_number in enumerate(pps_numbers):
        if i == 0:  #: # if first pps ping, ignore
            # TOTAL_SAMPLES starts at zero check
            assert (
                pps_table.loc[pps_numbers[i], "TOTAL_SAMPLES"] == 0
            ), "TOTAL_SAMPLES was not zero in first row of PPS file!"
            continue
        if pd.isnull(pps_table.loc[pps_numbers[i], "LAST_RMC_AUDIOMOTH_TIME"]):
            # no RMC data in this row
            continue

        # pps_table contains a row for each pps ping recieved, written immediately when the pps is recieved
        # the SAMPLES_WRITTEN column contains the current number of samples recorded to the WAV file
        #
        # The RMC data packet with GPS information relevant to a PPS arrives after the PPS
        # A row contains the _most recent_ RMC data. So if we are making one row per second and getting data
        # as expected, the RMC info will be in the row _after_ the PPS ping row to which it corresponds.
        # The RMC data includes the audiomoth time at which the RMC was recieved, and should be after the
        # pervious row's PPS Audiomoth time but before the current row's PPS Audiomoth time.
        #
        # let's call this row row1 and the previous row row0
        # check that LAST_RMC_AUDIOMOTH_TIME falls between AUDIOMOTH_TIME of row0 and row1
        # which are the times when the PPS ping was recieved, according to the AudioMoth clock.
        # If true, the RMC data in row1 corresponds to PPS ping in row0
        # So we can pair SAMPLES_WRITTEN from row0 with LAST_RMC_GPS_TIME from row1 to know
        # the number samples taken at an accurate timestamp
        row1 = pps_table.loc[pps_number]
        row0 = pps_table.loc[pps_numbers[i - 1]]

        # GPS Status Check
        # The GPS status should be "A" if it is synced properly
        # if not, discard this RMC info and go to the next row
        if row1["STATUS"] != "A":
            continue

        # GPS_TIME integer check: GPS_TIME should contain even seconds
        # if not, eg 3.301, something is wrong. Discard this RMC info and go to the next row
        rmc_gps_time = parse_RMC_time_to_seconds(row1["LAST_RMC_GPS_TIME"])
        if not np.isclose(rmc_gps_time % 1, 0, 1e-3):
            # print(f'rmc time was non-integer: pps: {pps_number}, gps time {rmc_gps_time}')
            continue  # non-integer RMC GPS TIME!

        # parse the timestamp of the LAST_RMC_AUDIOMOTH_TIME from this row, row1
        row1_last_rmc_amtime = parse_RMC_time_to_seconds(
            row1["LAST_RMC_AUDIOMOTH_TIME"]
        )
        # parse the timestamp of AUDIOMOTH_TIME (time PPS was recieved, internal clock) for previous row, row0
        row0_pps_amtime = parse_RMC_time_to_seconds(row0["AUDIOMOTH_TIME"])
        # parse the timestamp of AUDIOMOTH_TIME (time PPS was recieved, internal clock) for this row, row1
        # row1_pps_amtime  = parse_RMC_time_to_seconds(row1["AUDIOMOTH_TIME"])

        # RMC within one second of previous PPS check:
        # check that the RMC packet for this row was recieved within one second of the previous row's PPS
        # if so, we match the RMC packet's GPS timestamp with the previous PPS ping's info about samples recorded
        # we are using the internal AudioMoth clock to check the relative arrival time of messages, no GPS time is used yet
        # Note that sometimes the RMC packet for PPS 4 arrives (eg 0.3 seconds) after the arrival of PPS 5, so this condition can
        # be true even if the pairing is not correct. "observed smaple rate check" below attempts to catch these scenarios.
        if (
            row1_last_rmc_amtime > row0_pps_amtime
            and row1_last_rmc_amtime < row0_pps_amtime + 1.0
        ):
            # observed sample rate check
            # check observed sample rate since the last RMC + PPS pairing, if off by more than tolerance,
            # we assume this RMC is not aligned with the correct PPS
            # for example, the RMC message for PPS 4 may have arrived after PPS 5, so we would be attemping to pair
            # row1's RMC GPS time with the incorrect PPS information (SAMPLES_WRITTEN)
            # can only check if we have one pairing already, so skip the first row
            if len(samples_written) > 0:
                observed_sr = (row0["TOTAL_SAMPLES"] - samples_written[-1]) / (
                    rmc_gps_time - gps_times_sec[-1]
                )
                if abs(observed_sr - expected_sr) > sr_tol:
                    print(
                        f"Observed sr {observed_sr} did not match expected {expected_sr} at PPS {pps_number}. Discarding RMC."
                    )
                    continue

            # all of our checks have passed.
            # add the information to our lists of paired accurate time + samples take
            samples_written.append(row0["TOTAL_SAMPLES"])
            gps_times_str.append(row1["LAST_RMC_GPS_TIME"])
            pps_numbers_out.append(pps_numbers[i - 1])

            # todo: retain a timestamp rather than string
            gps_times_sec.append(rmc_gps_time)

            # also save the COUNTER (CPU clock cycle position, 0-999) from row0 (recorded when PPS is recieved)
            counter_positions.append(row0["COUNTER"])
        else:
            # this row's RMC data does not correspond to the previous row's PPS ping
            # so we don't add it to our table of accurate time and number of samples
            pps_pings_with_no_RMC.append(pps_number)

    print(
        f"There were {len(pps_pings_with_no_RMC)} PPS pings without corresponding RMC data in {pps_file}"
    )
    return pd.DataFrame(
        zip(
            pps_numbers_out,
            samples_written,
            gps_times_sec,
            gps_times_str,
            counter_positions,
        ),
        columns=[
            "PPS",
            "SAMPLES_WRITTEN",
            "GPS_TIME_SEC",
            "GPS_TIME_STR",
            "CPU_CLOCK_COUNTER",
        ],
    )


def correct_sample_rate(
    audio,
    processed_pps_df,
    desired_sr,
    expected_sr=48000,
    interpolation_method="nearest",
):
    """
    Args:
        - audio: Audio objecte
        - processed_pps_df: a dataframe of samples written and the corresponding RMC time.
                            2 columns:  SAMPLES_WRITTEN and RMC_TIME
        - desired_sr:   the desired sample_rate for the files.
        - expected_sr: the expected sample rate of the input;
            gives warning if input sr differs by more than 100 Hz on any interval
        - interpolation_method: Can be e.g. "linear", "quadratic", "nearest". See scipy.interpolate.interp1d docs.
            nearest is slightly faster than linear
    Returns:
        - Audio object with desired sample rate and updated .metadata['recording_start_time']
    Also writes out a .wav file with corrected sample_rate
    Each error in sampling rate is corrected between every RMC data
        if pps pings are received every second and desired_sr = 48000
        the first second is corrected to be exactly 48000, the second second is corrected to be
        exactly 48000 samples. These are concatenated to form the entire audio file.
    """

    dtype = audio.samples.dtype
    new_samples = []  # make container for the resampled samples

    for i in processed_pps_df.index[1:]:
        # don't do first row - there are 0 samples written so far!
        # find start and stop sample indices in the WAV file taken between RMC_TIME of the previous and current row
        # the samples we want are [previous row SAMPLES WRITTEN, this row SAMPLES_WRITTEN)
        # ie [include, exclude); python's `list[a:b]` syntax provides the desired slice [a,b)
        this_row = processed_pps_df.loc[i]
        previous_row = processed_pps_df.loc[i - 1]
        end_sample_exclude = this_row["SAMPLES_WRITTEN"]
        start_sample = previous_row["SAMPLES_WRITTEN"]

        time_elapsed = this_row["GPS_TIME_SEC"] - previous_row["GPS_TIME_SEC"]
        observed_sr = (end_sample_exclude - start_sample) / time_elapsed

        # check that the observed_sr is close to the desired_sr
        if abs(observed_sr - expected_sr) > 100:
            print(
                f"Warning: observed_sr = {observed_sr} is more than 100 Hz away from desired_sr = {expected_sr}. \
                  This may indicate a problem like missing audio data, or an error in the timestamps in the pps file. \
                  The offending line number in preprocessed_pps_df: {i}; PPS number {this_row['PPS']}"
            )

        ##interpolate samples to achieve desired_sr across the time interval
        raw_samples = audio.samples[start_sample:end_sample_exclude]
        # TODO try including 1 before and 1 after second period, so we can better interpolate?
        x_values = (
            np.arange(len(raw_samples)) / observed_sr
        )  # time in seconds of each sample, assuming samples were taken at a constant rate (even spacing)
        interpolator = interp1d(
            x_values, raw_samples, kind=interpolation_method, fill_value="extrapolate"
        )  # need to extrapolate for the points outside the original data (e.g. beginning and ends if upsampling)
        desired_x = (
            np.arange(int(time_elapsed * desired_sr)) / desired_sr
        )  # TODO do I need to check that this int rounding is OK. What if the rounding is adding some error??
        resampled = interpolator(desired_x)
        # resampled = resampled[0: int(desired_sr*time_elapsed)] #why would it have a different number of samples than this? why do we need to trim/subset?
        assert len(resampled) == int(time_elapsed * desired_sr)
        resampled = resampled.astype(dtype=dtype)

        new_samples.append(resampled)

    samples_out = np.concatenate(new_samples)
    # create new audio object with same metadata as the original
    new_audio = audio._spawn(samples=samples_out, sample_rate=desired_sr)
    # update the 'recording_start_time' in audio.metadata
    start_time = processed_pps_df.loc[0, "GPS_TIME_STR"]
    if new_audio.metadata is None:
        new_audio.metadata = {}
    new_audio.metadata["recording_start_time"] = parse_RMC_time_to_datetime(start_time)

    return new_audio


# timestamp rounding utilities


def time_mod(time, delta, epoch=None):
    if epoch is None:
        epoch = datetime(1970, 1, 1, tzinfo=time.tzinfo)
    return (time - epoch) % delta


def time_round(time, delta, epoch=None):
    mod = time_mod(time, delta, epoch)
    if mod < delta / 2:
        return time - mod
    return time + (delta - mod)


def time_floor(time, delta, epoch=None):
    mod = time_mod(time, delta, epoch)
    return time - mod


def time_ceil(time, delta, epoch=None):
    mod = time_mod(time, delta, epoch)
    if mod:
        return time + (delta - mod)
    return time


def trim_audio_by_datetime(
    audio, start_datetime, end_datetime=None, duration=None, out_of_bounds_error=True
):
    """trim Audio object using the embeded datetime in .metadata['recording_start_time']
    start_datetime and end_datetime should be _localized_ datetime objects, eg
    pytz.timezone('UTC').localize(datetime(2020,4,4,10,25,15))
    pass either end_datetime or duration
    out_of_bounds_error: if True, raise error for requesting before or after this recording's start/end
        If False, just use the part of the recording that overlaps with desired period
    """
    if duration is not None:
        assert end_datetime is None, "do not specify both duration and end_datetime"
        end_datetime = start_datetime + datetime.timedelta(seconds=duration)
    start_seconds = (
        start_datetime - audio.metadata["recording_start_time"]
    ).total_seconds()
    if start_seconds < 0:
        if out_of_bounds_error:
            raise ValueError(
                f"start_datetime was before the beginning of the Audio file ({audio.metadata['recording_start_time']})"
            )
        # audio.trim will just get the audio starting at time 0 (equivalent to start_seconds = 0)

    end_seconds = (
        end_datetime - audio.metadata["recording_start_time"]
    ).total_seconds()

    if end_seconds > audio.duration:
        if out_of_bounds_error:
            raise ValueError(
                f"The targeted end time was after the end of the Audio file"
            )
        # else, audio.trim will just get the audio until the end of the file

    return audio.trim(start_seconds, end_seconds)


def trim_to_start_on_minute(audio):
    ts = audio.metadata["recording_start_time"]
    if ts.second == 0:
        return audio
    else:
        start_at_next = timedelta(minutes=1)
        # new_ts = time_ceil(ts,start_at_next) #unused, but leaving for reference
        mod = time_mod(ts, start_at_next)
        offset = start_at_next.total_seconds() - mod.total_seconds()
        # audio.trim also updates the recording start time meatadata field
        return audio.trim(offset, audio.duration)


def sync_entire_sd(sd_folder, output_folder, output_sr, skip_if_completed=True):
    audio_files = list(sd_folder.glob("*.WAV"))
    success_record = []

    # make the directory for each sd card in the output folder
    out_sd = output_folder / sd_folder.stem
    out_sd.mkdir(exist_ok=True, parents=True)
    print(f"There are {len(audio_files)} files in {sd_folder}")

    for file in audio_files:
        # check if already completed
        out_filename = out_sd / file.name
        if out_filename.exists() and skip_if_completed:
            # check if file has proper metadata (if not, re-do it)
            try:
                p = parse_audiomoth_metadata_from_path(out_filename)
                assert "artist" in p
                # success -> skip
                success_record.append(True)
                continue
            except Exception as e:
                # If there's an error with the existing file, we'll reprocess it
                out_filename.unlink()  # Remove the problematic file

        try:
            # Get the processed PPS DF
            pps_file = file.parent / str(file.stem + ".PPS")
            assert pps_file.exists()
            processed_pps_df = process_pps_file(pps_file)

            # Resample the audio
            resampled_audio = correct_sample_rate(
                Audio.from_file(file, dtype=np.float16),
                processed_pps_df,
                desired_sr=output_sr,
            )

            # if recording starts with seconds!=0 in timestamp, trim to start at next even minute
            # this ensures the clips will align for 3s or 5s prediction windows
            resampled_audio = trim_to_start_on_minute(resampled_audio)

            # save
            processed_pps_df.to_csv(out_sd / file.with_suffix(".csv").name)
            resampled_audio.save(out_filename)
            success_record.append(True)

            # clean up memory after each file
            del resampled_audio
            gc.collect()
        except Exception as exc:
            print(f"Exception was raised while attempting to resample/sync {file}:")
            print(exc)
            success_record.append(False)
        finally:
            gc.collect()  # free memory after each file

    return pd.DataFrame(
        zip(audio_files, success_record), columns=["audio_file", "success"]
    )

	"""
	use the PPS data generated by AudioMothGPSDeploy_1_0_8_Hardware_1_1
	to generate audio files starting at
	a known time and having the exact desired sampling rate
	"""

	import pandas as pd
	import numpy as np
	from scipy.interpolate import interp1d
	from datetime import datetime, timedelta
	import pytz
	from opensoundscape import Audio
	import gc
	from aru_metadata_parser.parse import parse_audiomoth_metadata_from_path


	def parse_RMC_time_to_seconds(timestamp: str) -> float:
	"""
	Takes in the GPS-audiomoth formatted timestamp and returns unix epoch time in seconds.
	Args:
	timestamp: string in the format used by our GPS-audiomoths
	'%Y-%m-%dT%H:%M:%S.%f' e.g. "2022-05-14T10:30:00.725"
	Returns: time from unix-epoch in seconds (to 10^-6s resolution)as a float
	"""
	stamp_format = "%Y-%m-%dT%H:%M:%S.%f"
	return datetime.strptime(timestamp, stamp_format).timestamp()


	def parse_RMC_time_to_datetime(timestamp: str) -> datetime:
	"""
	Takes in the GPS-audiomoth formatted timestamp and returns datetime.datetime
	localized with UTC timezone. Assumes the input time string is given in UTC.

	Args:
	timestamp: string in the format used by our GPS-audiomoths
	'%Y-%m-%dT%H:%M:%S.%f' e.g. "2022-05-14T10:30:00.725"
	Returns: datetime object localized to UTC timezone
	"""
	stamp_format = "%Y-%m-%dT%H:%M:%S.%f"
	tz = pytz.timezone("UTC")
	return tz.localize(datetime.strptime(timestamp, stamp_format))


	def process_pps_file(pps_file, expected_sr=48000, sr_tol=50) -> pd.DataFrame:
	"""
	Using the PPS File, creates a table aligning GPS_TIME_SEC values with the corresponding PPS and SAMPLES_WRITTEN Values

	This is the information we need to interpolate audio files to achieve "synchronized" files with the exact desired sampling rate

	This function performs several checks to try to determine whether a value in the LAST_RMC_GPS_TIME column corresponds to the
	PPS from the previous line. If so, the previous line's SAMPLES_WRITTEN value gives the number of samples written at the time of
	the PPS, whose accurate time stamp is given by the LAST_RMC_GPS_TIME of the subsequent row.

	If any of the validation checks fail, we discard the RMC data from that row. As long as there is eventually a good pairing of
	RMC data (LAST_RMC_GPS_TIME) with a PPS, we will know the number of samples written and the exact time interval. This allows us
	to interpolate to achieve the nominal sampling rate.

	Args:
	- pps_file: path to the pps file (comma separated file produced by the firmware `AudioMothGPSDeploy_1_0_8_Hardware_1_1`)
	Returns:
	- a DataFrame with rows in which GPS_TIME_SEC is matched with the corresponding PPS, SAMPLES_WRITTEN, and COUNTER values
	columns: ["PPS","SAMPLES_WRITTEN", "GPS_TIME_SEC", "GPS_TIME_STR", "CPU_CLOCK_COUNTER"]

	"""
	pps_table = pd.read_csv(pps_file, index_col=0)

	samples_written = []
	gps_times_str = []
	gps_times_sec = []
	pps_numbers_out = []
	counter_positions = []

	pps_pings_with_no_RMC = []

	# Buffer overflow check: Fail if any buffer overflows occurred
	# a buffer overflow can be detected by a permanent difference of a multiple of 8 between
	# BUFFERS_FILLED and BUFFERS_WRITTEN. A difference of 8 corresponds to 2^17 samples that were
	# recorded from the sensor but never written to the WAV file. Thus TOTAL_SAMPLES is incorrect,
	# it overstates the number of samples actually written to the WAV file (it accurately reports the
	# number of samples recorded from the sensor into the buffer).
	# It may be possible to correct for buffer overflows by inserting 0s at the appropriate positions
	# in the WAV file, but here we fail rather than trying to correct buffer overflows.
	final_row = pps_table.iloc[-1]
	assert (
	final_row["BUFFERS_FILLED"] - final_row["BUFFERS_WRITTEN"] < 8
	), "Buffer overflow detected, WAV file has missing content"

	pps_numbers = pps_table.index

	# PPS Duplicate Check: check for duplicates in PPS number
	if sum(pps_numbers.duplicated()) > 0:
	raise ValueError(
	f"PPS numbers are duplicated in this file. Something may be wrong: {pps_file}"
	)

	# PPS Sequence Check: check that PPS numbers are sequential from first to last
	assert np.array_equal(
	pps_numbers.values, list(range(pps_numbers[0], pps_numbers[-1] + 1))
	), "PPS Numbers were not sequential integers"

	for i, pps_number in enumerate(pps_numbers):
	if i == 0: #: # if first pps ping, ignore
	# TOTAL_SAMPLES starts at zero check
	assert (
	pps_table.loc[pps_numbers[i], "TOTAL_SAMPLES"] == 0
	), "TOTAL_SAMPLES was not zero in first row of PPS file!"
	continue
	if pd.isnull(pps_table.loc[pps_numbers[i], "LAST_RMC_AUDIOMOTH_TIME"]):
	# no RMC data in this row
	continue

	# pps_table contains a row for each pps ping recieved, written immediately when the pps is recieved
	# the SAMPLES_WRITTEN column contains the current number of samples recorded to the WAV file
	#
	# The RMC data packet with GPS information relevant to a PPS arrives after the PPS
	# A row contains the _most recent_ RMC data. So if we are making one row per second and getting data
	# as expected, the RMC info will be in the row _after_ the PPS ping row to which it corresponds.
	# The RMC data includes the audiomoth time at which the RMC was recieved, and should be after the
	# pervious row's PPS Audiomoth time but before the current row's PPS Audiomoth time.
	#
	# let's call this row row1 and the previous row row0
	# check that LAST_RMC_AUDIOMOTH_TIME falls between AUDIOMOTH_TIME of row0 and row1
	# which are the times when the PPS ping was recieved, according to the AudioMoth clock.
	# If true, the RMC data in row1 corresponds to PPS ping in row0
	# So we can pair SAMPLES_WRITTEN from row0 with LAST_RMC_GPS_TIME from row1 to know
	# the number samples taken at an accurate timestamp
	row1 = pps_table.loc[pps_number]
	row0 = pps_table.loc[pps_numbers[i - 1]]

	# GPS Status Check
	# The GPS status should be "A" if it is synced properly
	# if not, discard this RMC info and go to the next row
	if row1["STATUS"] != "A":
	continue

	# GPS_TIME integer check: GPS_TIME should contain even seconds
	# if not, eg 3.301, something is wrong. Discard this RMC info and go to the next row
	rmc_gps_time = parse_RMC_time_to_seconds(row1["LAST_RMC_GPS_TIME"])
	if not np.isclose(rmc_gps_time % 1, 0, 1e-3):
	# print(f'rmc time was non-integer: pps: {pps_number}, gps time {rmc_gps_time}')
	continue # non-integer RMC GPS TIME!

	# parse the timestamp of the LAST_RMC_AUDIOMOTH_TIME from this row, row1
	row1_last_rmc_amtime = parse_RMC_time_to_seconds(
	row1["LAST_RMC_AUDIOMOTH_TIME"]
	)
	# parse the timestamp of AUDIOMOTH_TIME (time PPS was recieved, internal clock) for previous row, row0
	row0_pps_amtime = parse_RMC_time_to_seconds(row0["AUDIOMOTH_TIME"])
	# parse the timestamp of AUDIOMOTH_TIME (time PPS was recieved, internal clock) for this row, row1
	# row1_pps_amtime = parse_RMC_time_to_seconds(row1["AUDIOMOTH_TIME"])

	# RMC within one second of previous PPS check:
	# check that the RMC packet for this row was recieved within one second of the previous row's PPS
	# if so, we match the RMC packet's GPS timestamp with the previous PPS ping's info about samples recorded
	# we are using the internal AudioMoth clock to check the relative arrival time of messages, no GPS time is used yet
	# Note that sometimes the RMC packet for PPS 4 arrives (eg 0.3 seconds) after the arrival of PPS 5, so this condition can
	# be true even if the pairing is not correct. "observed smaple rate check" below attempts to catch these scenarios.
	if (
	row1_last_rmc_amtime > row0_pps_amtime
	and row1_last_rmc_amtime < row0_pps_amtime + 1.0
	):
	# observed sample rate check
	# check observed sample rate since the last RMC + PPS pairing, if off by more than tolerance,
	# we assume this RMC is not aligned with the correct PPS
	# for example, the RMC message for PPS 4 may have arrived after PPS 5, so we would be attemping to pair
	# row1's RMC GPS time with the incorrect PPS information (SAMPLES_WRITTEN)
	# can only check if we have one pairing already, so skip the first row
	if len(samples_written) > 0:
	observed_sr = (row0["TOTAL_SAMPLES"] - samples_written[-1]) / (
	rmc_gps_time - gps_times_sec[-1]
	)
	if abs(observed_sr - expected_sr) > sr_tol:
	print(
	f"Observed sr {observed_sr} did not match expected {expected_sr} at PPS {pps_number}. Discarding RMC."
	)
	continue

	# all of our checks have passed.
	# add the information to our lists of paired accurate time + samples take
	samples_written.append(row0["TOTAL_SAMPLES"])
	gps_times_str.append(row1["LAST_RMC_GPS_TIME"])
	pps_numbers_out.append(pps_numbers[i - 1])

	# todo: retain a timestamp rather than string
	gps_times_sec.append(rmc_gps_time)

	# also save the COUNTER (CPU clock cycle position, 0-999) from row0 (recorded when PPS is recieved)
	counter_positions.append(row0["COUNTER"])
	else:
	# this row's RMC data does not correspond to the previous row's PPS ping
	# so we don't add it to our table of accurate time and number of samples
	pps_pings_with_no_RMC.append(pps_number)

	print(
	f"There were {len(pps_pings_with_no_RMC)} PPS pings without corresponding RMC data in {pps_file}"
	)
	return pd.DataFrame(
	zip(
	pps_numbers_out,
	samples_written,
	gps_times_sec,
	gps_times_str,
	counter_positions,
	),
	columns=[
	"PPS",
	"SAMPLES_WRITTEN",
	"GPS_TIME_SEC",
	"GPS_TIME_STR",
	"CPU_CLOCK_COUNTER",
	],
	)


	def correct_sample_rate(
	audio,
	processed_pps_df,
	desired_sr,
	expected_sr=48000,
	interpolation_method="nearest",
	):
	"""
	Args:
	- audio: Audio objecte
	- processed_pps_df: a dataframe of samples written and the corresponding RMC time.
	2 columns: SAMPLES_WRITTEN and RMC_TIME
	- desired_sr: the desired sample_rate for the files.
	- expected_sr: the expected sample rate of the input;
	gives warning if input sr differs by more than 100 Hz on any interval
	- interpolation_method: Can be e.g. "linear", "quadratic", "nearest". See scipy.interpolate.interp1d docs.
	nearest is slightly faster than linear
	Returns:
	- Audio object with desired sample rate and updated .metadata['recording_start_time']

	Also writes out a .wav file with corrected sample_rate
	Each error in sampling rate is corrected between every RMC data
	if pps pings are received every second and desired_sr = 48000
	the first second is corrected to be exactly 48000, the second second is corrected to be
	exactly 48000 samples. These are concatenated to form the entire audio file.
	"""

	dtype = audio.samples.dtype
	new_samples = [] # make container for the resampled samples

	for i in processed_pps_df.index[1:]:
	# don't do first row - there are 0 samples written so far!
	# find start and stop sample indices in the WAV file taken between RMC_TIME of the previous and current row
	# the samples we want are [previous row SAMPLES WRITTEN, this row SAMPLES_WRITTEN)
	# ie [include, exclude); python's `list[a:b]` syntax provides the desired slice [a,b)
	this_row = processed_pps_df.loc[i]
	previous_row = processed_pps_df.loc[i - 1]
	end_sample_exclude = this_row["SAMPLES_WRITTEN"]
	start_sample = previous_row["SAMPLES_WRITTEN"]

	time_elapsed = this_row["GPS_TIME_SEC"] - previous_row["GPS_TIME_SEC"]
	observed_sr = (end_sample_exclude - start_sample) / time_elapsed

	# check that the observed_sr is close to the desired_sr
	if abs(observed_sr - expected_sr) > 100:
	print(
	f"Warning: observed_sr = {observed_sr} is more than 100 Hz away from desired_sr = {expected_sr}. \
	This may indicate a problem like missing audio data, or an error in the timestamps in the pps file. \
	The offending line number in preprocessed_pps_df: {i}; PPS number {this_row['PPS']}"
	)

	##interpolate samples to achieve desired_sr across the time interval
	raw_samples = audio.samples[start_sample:end_sample_exclude]
	# TODO try including 1 before and 1 after second period, so we can better interpolate?
	x_values = (
	np.arange(len(raw_samples)) / observed_sr
	) # time in seconds of each sample, assuming samples were taken at a constant rate (even spacing)
	interpolator = interp1d(
	x_values, raw_samples, kind=interpolation_method, fill_value="extrapolate"
	) # need to extrapolate for the points outside the original data (e.g. beginning and ends if upsampling)
	desired_x = (
	np.arange(int(time_elapsed * desired_sr)) / desired_sr
	) # TODO do I need to check that this int rounding is OK. What if the rounding is adding some error??
	resampled = interpolator(desired_x)
	# resampled = resampled[0: int(desired_sr*time_elapsed)] #why would it have a different number of samples than this? why do we need to trim/subset?
	assert len(resampled) == int(time_elapsed * desired_sr)
	resampled = resampled.astype(dtype=dtype)

	new_samples.append(resampled)

	samples_out = np.concatenate(new_samples)
	# create new audio object with same metadata as the original
	new_audio = audio._spawn(samples=samples_out, sample_rate=desired_sr)
	# update the 'recording_start_time' in audio.metadata
	start_time = processed_pps_df.loc[0, "GPS_TIME_STR"]
	if new_audio.metadata is None:
	new_audio.metadata = {}
	new_audio.metadata["recording_start_time"] = parse_RMC_time_to_datetime(start_time)

	return new_audio


	# timestamp rounding utilities


	def time_mod(time, delta, epoch=None):
	if epoch is None:
	epoch = datetime(1970, 1, 1, tzinfo=time.tzinfo)
	return (time - epoch) % delta


	def time_round(time, delta, epoch=None):
	mod = time_mod(time, delta, epoch)
	if mod < delta / 2:
	return time - mod
	return time + (delta - mod)


	def time_floor(time, delta, epoch=None):
	mod = time_mod(time, delta, epoch)
	return time - mod


	def time_ceil(time, delta, epoch=None):
	mod = time_mod(time, delta, epoch)
	if mod:
	return time + (delta - mod)
	return time


	def trim_audio_by_datetime(
	audio, start_datetime, end_datetime=None, duration=None, out_of_bounds_error=True
	):
	"""trim Audio object using the embeded datetime in .metadata['recording_start_time']

	start_datetime and end_datetime should be _localized_ datetime objects, eg
	pytz.timezone('UTC').localize(datetime(2020,4,4,10,25,15))

	pass either end_datetime or duration
	out_of_bounds_error: if True, raise error for requesting before or after this recording's start/end
	If False, just use the part of the recording that overlaps with desired period
	"""
	if duration is not None:
	assert end_datetime is None, "do not specify both duration and end_datetime"
	end_datetime = start_datetime + datetime.timedelta(seconds=duration)
	start_seconds = (
	start_datetime - audio.metadata["recording_start_time"]
	).total_seconds()
	if start_seconds < 0:
	if out_of_bounds_error:
	raise ValueError(
	f"start_datetime was before the beginning of the Audio file ({audio.metadata['recording_start_time']})"
	)
	# audio.trim will just get the audio starting at time 0 (equivalent to start_seconds = 0)

	end_seconds = (
	end_datetime - audio.metadata["recording_start_time"]
	).total_seconds()

	if end_seconds > audio.duration:
	if out_of_bounds_error:
	raise ValueError(
	f"The targeted end time was after the end of the Audio file"
	)
	# else, audio.trim will just get the audio until the end of the file

	return audio.trim(start_seconds, end_seconds)


	def trim_to_start_on_minute(audio):
	ts = audio.metadata["recording_start_time"]
	if ts.second == 0:
	return audio
	else:
	start_at_next = timedelta(minutes=1)
	# new_ts = time_ceil(ts,start_at_next) #unused, but leaving for reference
	mod = time_mod(ts, start_at_next)
	offset = start_at_next.total_seconds() - mod.total_seconds()
	# audio.trim also updates the recording start time meatadata field
	return audio.trim(offset, audio.duration)


	def sync_entire_sd(sd_folder, output_folder, output_sr, skip_if_completed=True):
	audio_files = list(sd_folder.glob("*.WAV"))
	success_record = []

	# make the directory for each sd card in the output folder
	out_sd = output_folder / sd_folder.stem
	out_sd.mkdir(exist_ok=True, parents=True)
	print(f"There are {len(audio_files)} files in {sd_folder}")

	for file in audio_files:
	# check if already completed
	out_filename = out_sd / file.name
	if out_filename.exists() and skip_if_completed:
	# check if file has proper metadata (if not, re-do it)
	try:
	p = parse_audiomoth_metadata_from_path(out_filename)
	assert "artist" in p
	# success -> skip
	success_record.append(True)
	continue
	except Exception as e:
	# If there's an error with the existing file, we'll reprocess it
	out_filename.unlink() # Remove the problematic file

	try:
	# Get the processed PPS DF
	pps_file = file.parent / str(file.stem + ".PPS")
	assert pps_file.exists()
	processed_pps_df = process_pps_file(pps_file)

	# Resample the audio
	resampled_audio = correct_sample_rate(
	Audio.from_file(file, dtype=np.float16),
	processed_pps_df,
	desired_sr=output_sr,
	)

	# if recording starts with seconds!=0 in timestamp, trim to start at next even minute
	# this ensures the clips will align for 3s or 5s prediction windows
	resampled_audio = trim_to_start_on_minute(resampled_audio)

	# save
	processed_pps_df.to_csv(out_sd / file.with_suffix(".csv").name)
	resampled_audio.save(out_filename)
	success_record.append(True)

	# clean up memory after each file
	del resampled_audio
	gc.collect()
	except Exception as exc:
	print(f"Exception was raised while attempting to resample/sync {file}:")
	print(exc)
	success_record.append(False)
	finally:
	gc.collect() # free memory after each file

	return pd.DataFrame(
	zip(audio_files, success_record), columns=["audio_file", "success"]
	)