diff --git a/.gitignore b/.gitignore
index 8c81e19..7b1d036 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,6 +1,7 @@
 models
-data
 dependencies
+data/test*
+data/export
 
 .idea
 
diff --git a/align/export.py b/align/export.py
new file mode 100644
index 0000000..b2ca92c
--- /dev/null
+++ b/align/export.py
@@ -0,0 +1,362 @@
+import os
+import sys
+import csv
+import math
+import json
+import random
+import logging
+import argparse
+import statistics
+import os.path as path
+
+from tqdm import tqdm
+from pydub import AudioSegment
+from collections import Counter
+from multiprocessing import Pool
+
+unknown = '<unknown>'
+
+
+def fail(message, code=1):
+    logging.fatal(message)
+    exit(code)
+
+
+def engroup(lst, get_key):
+    groups = {}
+    for obj in lst:
+        key = get_key(obj)
+        if key in groups:
+            groups[key].append(obj)
+        else:
+            groups[key] = [obj]
+    return groups
+
+
+def get_set_sizes(population_size):
+    margin_of_error = 0.01
+    fraction_picking = 0.50
+    z_score = 2.58  # Corresponds to confidence level 99%
+    numerator = (z_score ** 2 * fraction_picking * (1 - fraction_picking)) / (
+            margin_of_error ** 2
+    )
+    sample_size = 0
+    for train_size in range(population_size, 0, -1):
+        denominator = 1 + (z_score ** 2 * fraction_picking * (1 - fraction_picking)) / (
+                margin_of_error ** 2 * train_size
+        )
+        sample_size = int(numerator / denominator)
+        if 2 * sample_size + train_size <= population_size:
+            break
+    return population_size - 2 * sample_size, sample_size
+
+
+def load_segment(audio_path):
+    return audio_path, AudioSegment.from_file(audio_path)
+
+
+def load_segment_dry(audio_path):
+    if path.isfile(audio_path):
+        logging.debug('Would load file "{}"'.format(audio_path))
+    else:
+        fail('File not found: "{}"'.format(audio_path))
+    return audio_path, AudioSegment.empty()
+
+
+def main(args):
+    parser = argparse.ArgumentParser(description='Export aligned speech samples.')
+
+    parser.add_argument('--audio', type=str,
+                        help='Take audio file as input (requires "--aligned <file>")')
+    parser.add_argument('--aligned', type=str,
+                        help='Take alignment file ("<...>.aligned") as input (requires "--audio <file>")')
+    parser.add_argument('--catalog', type=str,
+                        help='Take alignment and audio file references of provided catalog ("<...>.catalog") as input')
+    parser.add_argument('--ignore-missing', action="store_true",
+                        help='Ignores catalog entries with missing files')
+    parser.add_argument('--target-dir', type=str, required=True,
+                        help='Existing target directory for storing generated sets (files and directories)')
+    parser.add_argument('--filter', type=str,
+                        help='Python expression that computes a boolean value from sample data fields. '
+                             'If the result is True, the sample will be dropped.')
+    parser.add_argument('--criteria', type=str, default='100',
+                        help='Python expression that computes a number as quality indicator from sample data fields.')
+    parser.add_argument('--partition', type=str, action='append',
+                        help='Expression of the form "<number>:<partition>" where all samples with a quality indicator '
+                             '(--criteria) above or equal the given number and below the next bigger one are assigned '
+                             'to the specified partition. Samples below the lowest partition criteria are assigned to '
+                             'partition "other".')
+    parser.add_argument('--split', action="store_true",
+                        help='Split each partition except "other" into train/dev/test sub-sets.')
+    parser.add_argument('--split-field', type=str,
+                        help='Sample meta field that should be used for splitting (e.g. "speaker")')
+    parser.add_argument('--split-seed', type=int,
+                        help='Random seed for set splitting')
+    parser.add_argument('--debias', type=str, action='append',
+                        help='Sample meta field to group samples for debiasing (e.g. "speaker"). '
+                             'Group sizes will be capped according to --debias-sigma-factor')
+    parser.add_argument('--debias-sigma-factor', type=float, default=3.0,
+                        help='Standard deviation (sigma) factor after which the sample number of a group gets capped')
+    parser.add_argument('--loglevel', type=int, default=20,
+                        help='Log level (between 0 and 50) - default: 20')
+    parser.add_argument('--no-progress', action="store_true",
+                        help='Prevents showing progress bars')
+    parser.add_argument('--format', type=str, default='csv',
+                        help='Sample list format - one of (json|csv)')
+    parser.add_argument('--rate', type=int,
+                        help='Export wav-files with this sample rate')
+    parser.add_argument('--channels', type=int,
+                        help='Export wav-files with this number of channels')
+    parser.add_argument('--force', action="store_true",
+                        help='Overwrite existing files')
+    parser.add_argument('--dry-run', action="store_true",
+                        help='Simulates export without writing or creating any file or directory')
+    parser.add_argument('--dry-run-fast', action="store_true",
+                        help='Simulates export without writing or creating any file or directory. '
+                             'In contrast to --dry-run this faster simulation will not load samples.')
+    parser.add_argument('--workers', type=int, default=None,
+                        help='Number of workers for loading and re-sampling audio files. Default: Number of CPUs')
+    parser.add_argument('--pretty', action="store_true",
+                        help='Writes indented JSON output')
+
+    args = parser.parse_args()
+
+    logging.basicConfig(stream=sys.stderr, level=args.loglevel if args.loglevel else 20)
+
+    def progress(iter, desc='Processing', total=None):
+        desc = desc.rjust(24)
+        return iter if args.no_progress else tqdm(iter, desc=desc, total=total, ncols=120)
+
+    logging.debug("Start")
+
+    pairs = []
+
+    def check_path(target_path, fs_type='file'):
+        if not (path.isfile(target_path) if fs_type == 'file' else path.isdir(target_path)):
+            logging.fatal('{} not existing: "{}"'.format(fs_type[0].upper() + fs_type[1:], target_path))
+            exit(1)
+        return path.abspath(target_path)
+
+    def make_absolute(base_path, spec_path):
+        if not path.isabs(spec_path):
+            spec_path = path.join(base_path, spec_path)
+        spec_path = path.abspath(spec_path)
+        return spec_path if path.isfile(spec_path) else None
+
+    target_dir = check_path(args.target_dir, fs_type='directory')
+    if args.audio:
+        if args.aligned:
+            pairs.append((check_path(args.audio), check_path(args.aligned)))
+        else:
+            fail('If you specify "--audio", you also have to specify "--aligned"')
+    elif args.aligned:
+        fail('If you specify "--aligned", you also have to specify "--audio"')
+    elif args.catalog:
+        catalog = check_path(args.catalog)
+        catalog_dir = path.dirname(catalog)
+        with open(catalog, 'r') as catalog_file:
+            catalog_entries = json.load(catalog_file)
+        for entry in progress(catalog_entries, desc='Reading catalog'):
+            audio = make_absolute(catalog_dir, entry['audio'])
+            aligned = make_absolute(catalog_dir, entry['aligned'])
+            if audio is None or aligned is None:
+                if args.ignore_missing:
+                    continue
+                if audio is None:
+                    fail('Problem loading catalog "{}": Missing referenced audio file "{}"'
+                         .format(args.catalog, entry['audio']))
+                if aligned is None:
+                    fail('Problem loading catalog "{}": Missing referenced alignment file "{}"'
+                         .format(args.catalog, entry['aligned']))
+            pairs.append((audio, aligned))
+    else:
+        fail('You have to either specify "--audio" and "--aligned" or "--catalog"')
+
+    dry_run = args.dry_run or args.dry_run_fast
+    load_samples = not args.dry_run_fast
+
+    partition_specs = []
+    if args.partition is not None:
+        for partition_expr in args.partition:
+            parts = partition_expr.split(':')
+            if len(parts) != 2:
+                fail('Wrong partition specification: "{}"'.format(partition_expr))
+            partition_specs.append((float(parts[0]), str(parts[1])))
+    partition_specs.sort(key=lambda p: p[0], reverse=True)
+
+    fragments = []
+    for audio_path, aligned_path in progress(pairs, desc='Loading alignments'):
+        with open(aligned_path, 'r') as aligned_file:
+            aligned_fragments = json.load(aligned_file)
+        for fragment in aligned_fragments:
+            fragment['audio_path'] = audio_path
+            fragments.append(fragment)
+
+    if args.filter is not None:
+        kept_fragments = []
+        for fragment in progress(fragments, desc='Filtering'):
+            if not eval(args.filter, {'math': math}, fragment):
+                kept_fragments.append(fragment)
+        if len(kept_fragments) < len(fragments):
+            logging.info('Filtered out {} samples'.format(len(fragments) - len(kept_fragments)))
+        fragments = kept_fragments
+        if len(fragments) == 0:
+            fail('Filter left no samples samples')
+
+    for fragment in progress(fragments, desc='Computing qualities'):
+        fragment['quality'] = eval(args.criteria, {'math': math}, fragment)
+
+    def get_meta(fragment, meta_field):
+        if 'meta' in fragment:
+            meta = fragment['meta']
+            if meta_field in meta:
+                for value in meta[meta_field]:
+                    return value
+        return unknown
+
+    if args.debias is not None:
+        for debias in args.debias:
+            grouped = engroup(fragments, lambda f: get_meta(f, debias))
+            if unknown in grouped:
+                fragments = grouped[unknown]
+                del grouped[unknown]
+            else:
+                fragments = []
+            counts = list(map(lambda f: len(f), grouped.values()))
+            mean = statistics.mean(counts)
+            sigma = statistics.pstdev(counts, mu=mean)
+            cap = int(mean + args.debias_sigma_factor * sigma)
+            counter = Counter()
+            for group, values in progress(grouped.items(), desc='Debiasing "{}"'.format(debias)):
+                if len(values) > cap:
+                    values.sort(key=lambda g: g['quality'])
+                    counter[group] += len(values) - cap
+                    values = values[-cap:]
+                fragments.extend(values)
+            logging.info('Dropped for debiasing "{}":'.format(debias))
+            for group, count in counter.most_common():
+                logging.info(' - "{}": {}'.format(group, count))
+
+    def get_partition(f):
+        quality = f['quality']
+        for minimum, partition_name in partition_specs:
+            if quality >= minimum:
+                return partition_name
+        return 'other'
+
+    lists = {}
+
+    def ensure_list(name):
+        lists[name] = []
+        if not args.force:
+            for p in [name, name + '.' + args.format]:
+                if path.exists(path.join(target_dir, p)):
+                    fail('"{}" already existing - use --force to ignore'.format(p))
+
+    if args.split_seed is not None:
+        random.seed(args.split_seed)
+    partitions = engroup(fragments, get_partition)
+    for partition, partition_fragments in partitions.items():
+        logging.info('Partition "{}":'.format(partition))
+        if not args.split or partition == 'other':
+            ensure_list(partition)
+            for fragment in partition_fragments:
+                fragment['list-name'] = partition
+            logging.info(' - samples: {}'.format(len(partition_fragments)))
+        else:
+            train_size, sample_size = get_set_sizes(len(partition_fragments))
+            if args.split_field:
+                portions = engroup(partition_fragments, lambda f: get_meta(f, args.split_field)).values()
+                portions.sort(key=lambda p: len(p))
+                train_set, dev_set, test_set = [], [], []
+                for offset, sample_set in [(0, dev_set), (1, test_set)]:
+                    for portion in portions[offset::2]:
+                        if len(sample_set) < sample_size:
+                            sample_set.extend(portion)
+                        else:
+                            train_set.extend(portion)
+            else:
+                random.shuffle(partition_fragments)
+                test_set = partition_fragments[:sample_size]
+                partition_fragments = partition_fragments[sample_size:]
+                dev_set = partition_fragments[:sample_size]
+                train_set = partition_fragments[sample_size:]
+            for set_name, set_fragments in [('train', train_set), ('dev', dev_set), ('test', test_set)]:
+                list_name = partition + '-' + set_name
+                ensure_list(list_name)
+                for fragment in set_fragments:
+                    fragment['list-name'] = list_name
+                logging.info(' - sub-set "{}" - samples: {}'.format(set_name, len(set_fragments)))
+
+    for list_name in lists.keys():
+        dir_name = path.join(target_dir, list_name)
+        if not path.isdir(dir_name):
+            if dry_run:
+                logging.debug('Would create directory "{}"'.format(dir_name))
+            else:
+                os.mkdir(dir_name)
+
+    def list_fragments():
+        audio_files = engroup(fragments, lambda f: f['audio_path'])
+        pool = Pool(args.workers)
+        ls = load_segment if load_samples else load_segment_dry
+        for audio_path, audio in pool.imap_unordered(ls, audio_files.keys()):
+            file_fragments = audio_files[audio_path]
+            if args.channels is not None:
+                audio = audio.set_channels(args.channels)
+            if args.rate is not None:
+                audio = audio.set_frame_rate(args.rate)
+            file_fragments.sort(key=lambda f: f['start'])
+            for fragment in file_fragments:
+                if load_samples:
+                    yield audio[fragment['start']:fragment['end']], fragment
+                else:
+                    yield audio, fragment
+
+    for audio_segment, fragment in progress(list_fragments(), desc='Exporting samples', total=len(fragments)):
+        list_name = fragment['list-name']
+        group_list = lists[list_name]
+        sample_name = 'sample-{:010d}.wav'.format(len(group_list))
+        rel_path = path.join(list_name, sample_name)
+        abs_path = path.join(target_dir, rel_path)
+        if dry_run:
+            logging.debug('Would write file "{}"'.format(abs_path))
+        else:
+            with open(abs_path, "wb") as wav_file:
+                audio_segment.export(wav_file, format="wav")
+                file_size = wav_file.tell()
+            group_list.append((rel_path, file_size, fragment))
+
+    for list_name, group_list in progress(lists.items(), desc='Writing lists'):
+        if args.format == 'json':
+            json_path = path.join(target_dir, list_name + '.json')
+            if dry_run:
+                logging.debug('Would write file "{}"'.format(json_path))
+            else:
+                entries = []
+                for rel_path, file_size, fragment in group_list:
+                    entry = {
+                        'audio': rel_path,
+                        'size': file_size,
+                        'transcript': fragment['aligned'],
+                        'duration': fragment['end'] - fragment['start']
+                    }
+                    if 'aligned-raw' in fragment:
+                        entry['transcript-raw'] = fragment['aligned-raw']
+                    entries.append(entry)
+                with open(json_path, 'w') as json_file:
+                    json.dump(entries, json_file)
+        else:
+            csv_path = path.join(target_dir, list_name + '.csv')
+            if dry_run:
+                logging.debug('Would write file "{}"'.format(csv_path))
+            else:
+                with open(csv_path, 'w') as csv_file:
+                    writer = csv.writer(csv_file)
+                    for rel_path, file_size, fragment in group_list:
+                        writer.writerow([rel_path, file_size, fragment['aligned']])
+
+
+if __name__ == '__main__':
+    main(sys.argv[1:])
+    os.system('stty sane')
diff --git a/bin/export.sh b/bin/export.sh
new file mode 100755
index 0000000..56850cf
--- /dev/null
+++ b/bin/export.sh
@@ -0,0 +1,4 @@
+#!/usr/bin/env bash
+approot="$(dirname "$(dirname "$(readlink -fm "$0")")")"
+source "$approot/venv/bin/activate"
+python "$approot/align/export.py" "$@"
diff --git a/bin/statistics.sh b/bin/statistics.sh
new file mode 100755
index 0000000..57b2317
--- /dev/null
+++ b/bin/statistics.sh
@@ -0,0 +1,4 @@
+#!/usr/bin/env bash
+approot="$(dirname "$(dirname "$(readlink -fm "$0")")")"
+source "$approot/venv/bin/activate"
+python "$approot/align/statistics.py" "$@"
diff --git a/data/all.catalog b/data/all.catalog
new file mode 100644
index 0000000..c369ea2
--- /dev/null
+++ b/data/all.catalog
@@ -0,0 +1,12 @@
+[
+  {
+    "audio": "test1/joined.mp3",
+    "text": "test1/transcript.txt",
+    "aligned": "test1/aligned.json"
+  },
+  {
+    "audio": "test2/joined.mp3",
+    "text": "test2/transcript.script",
+    "aligned": "test2/aligned.json"
+  }
+]
diff --git a/requirements.txt b/requirements.txt
index 8b75cf0..878ef51 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -3,3 +3,4 @@ deepspeech
 webrtcvad
 tqdm
 textdistance
+pydub