metadata

pretty_name: AMI Corpus
annotations_creators:
  - expert-generated
language_creators:
  - crowdsourced
  - expert-generated
language:
  - en
license:
  - cc-by-4.0
multilinguality:
  - monolingual
size_categories:
  - 100K<n<1M
source_datasets:
  - original
task_categories:
  - automatic-speech-recognition
task_ids: []
dataset_info:
  - config_name: microphone-single
    features:
      - name: word_ids
        sequence: string
      - name: word_start_times
        sequence: float32
      - name: word_end_times
        sequence: float32
      - name: word_speakers
        sequence: string
      - name: segment_ids
        sequence: string
      - name: segment_start_times
        sequence: float32
      - name: segment_end_times
        sequence: float32
      - name: segment_speakers
        sequence: string
      - name: words
        sequence: string
      - name: channels
        sequence: string
      - name: file
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: train
        num_bytes: 42013753
        num_examples: 134
      - name: validation
        num_bytes: 5110497
        num_examples: 18
      - name: test
        num_bytes: 4821283
        num_examples: 16
    download_size: 11387715153
    dataset_size: 51945533
  - config_name: microphone-multi
    features:
      - name: word_ids
        sequence: string
      - name: word_start_times
        sequence: float32
      - name: word_end_times
        sequence: float32
      - name: word_speakers
        sequence: string
      - name: segment_ids
        sequence: string
      - name: segment_start_times
        sequence: float32
      - name: segment_end_times
        sequence: float32
      - name: segment_speakers
        sequence: string
      - name: words
        sequence: string
      - name: channels
        sequence: string
      - name: file-1-1
        dtype: string
      - name: file-1-2
        dtype: string
      - name: file-1-3
        dtype: string
      - name: file-1-4
        dtype: string
      - name: file-1-5
        dtype: string
      - name: file-1-6
        dtype: string
      - name: file-1-7
        dtype: string
      - name: file-1-8
        dtype: string
    splits:
      - name: train
        num_bytes: 42126341
        num_examples: 134
      - name: validation
        num_bytes: 5125645
        num_examples: 18
      - name: test
        num_bytes: 4834751
        num_examples: 16
    download_size: 90941506169
    dataset_size: 52086737
  - config_name: headset-single
    features:
      - name: word_ids
        sequence: string
      - name: word_start_times
        sequence: float32
      - name: word_end_times
        sequence: float32
      - name: word_speakers
        sequence: string
      - name: segment_ids
        sequence: string
      - name: segment_start_times
        sequence: float32
      - name: segment_end_times
        sequence: float32
      - name: segment_speakers
        sequence: string
      - name: words
        sequence: string
      - name: channels
        sequence: string
      - name: file
        dtype: string
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
    splits:
      - name: train
        num_bytes: 42491091
        num_examples: 136
      - name: validation
        num_bytes: 5110497
        num_examples: 18
      - name: test
        num_bytes: 4821283
        num_examples: 16
    download_size: 11505070978
    dataset_size: 52422871
  - config_name: headset-multi
    features:
      - name: word_ids
        sequence: string
      - name: word_start_times
        sequence: float32
      - name: word_end_times
        sequence: float32
      - name: word_speakers
        sequence: string
      - name: segment_ids
        sequence: string
      - name: segment_start_times
        sequence: float32
      - name: segment_end_times
        sequence: float32
      - name: segment_speakers
        sequence: string
      - name: words
        sequence: string
      - name: channels
        sequence: string
      - name: file-0
        dtype: string
      - name: file-1
        dtype: string
      - name: file-2
        dtype: string
      - name: file-3
        dtype: string
    splits:
      - name: train
        num_bytes: 42540063
        num_examples: 136
      - name: validation
        num_bytes: 5116989
        num_examples: 18
      - name: test
        num_bytes: 4827055
        num_examples: 16
    download_size: 45951596391
    dataset_size: 52484107

Dataset Card for AMI Corpus

Dataset Description
Dataset Structure
Dataset Creation
Considerations for Using the Data
Additional Information

Deprecated: This legacy dataset is outdated. Please, use edinburghcstr/ami instead.

Dataset Description

Homepage: AMI corpus
Repository: [Needs More Information]
Paper: [Needs More Information]
Leaderboard: [Needs More Information]
Point of Contact: [Needs More Information]

Dataset Summary

The AMI Meeting Corpus consists of 100 hours of meeting recordings. The recordings use a range of signals synchronized to a common timeline. These include close-talking and far-field microphones, individual and room-view video cameras, and output from a slide projector and an electronic whiteboard. During the meetings, the participants also have unsynchronized pens available to them that record what is written. The meetings were recorded in English using three different rooms with different acoustic properties, and include mostly non-native speakers.

Dataset Preprocessing

Individual samples of the AMI dataset contain very large audio files (between 10 and 60 minutes). Such lengths are unfeasible for most speech recognition models. In the following, we show how the dataset can effectively be chunked into multiple segments as defined by the dataset creators.

The following function cuts the long audio files into the defined segment lengths:

import librosa
import math
from datasets import load_dataset

SAMPLE_RATE = 16_000

def chunk_audio(batch):
    new_batch = {
        "audio": [],
        "words": [],
        "speaker": [],
        "lengths": [],
        "word_start_times": [],
        "segment_start_times": [],
    }

    audio, _ = librosa.load(batch["file"][0], sr=SAMPLE_RATE)

    word_idx = 0
    num_words = len(batch["words"][0])
    for segment_idx in range(len(batch["segment_start_times"][0])):
        words = []
        word_start_times = []
        start_time = batch["segment_start_times"][0][segment_idx]
        end_time = batch["segment_end_times"][0][segment_idx]

        # go back and forth with word_idx since segments overlap with each other
        while (word_idx > 1) and (start_time < batch["word_end_times"][0][word_idx - 1]):
            word_idx -= 1

        while word_idx < num_words and (start_time > batch["word_start_times"][0][word_idx]):
            word_idx += 1

        new_batch["audio"].append(audio[int(start_time * SAMPLE_RATE): int(end_time * SAMPLE_RATE)])

        while word_idx < num_words and batch["word_start_times"][0][word_idx] < end_time:
            words.append(batch["words"][0][word_idx])
            word_start_times.append(batch["word_start_times"][0][word_idx])
            word_idx += 1

        new_batch["lengths"].append(end_time - start_time)
        new_batch["words"].append(words)
        new_batch["speaker"].append(batch["segment_speakers"][0][segment_idx])
        new_batch["word_start_times"].append(word_start_times)

        new_batch["segment_start_times"].append(batch["segment_start_times"][0][segment_idx])

    return new_batch
    
ami = load_dataset("ami", "headset-single")
ami = ami.map(chunk_audio, batched=True, batch_size=1, remove_columns=ami["train"].column_names)

The segmented audio files can still be as long as a minute. To further chunk the data into shorter audio chunks, you can use the following script.

MAX_LENGTH_IN_SECONDS = 20.0

def chunk_into_max_n_seconds(batch):
    new_batch = {
        "audio": [],
        "text": [],
    }

    sample_length = batch["lengths"][0]
    segment_start = batch["segment_start_times"][0]

    if sample_length > MAX_LENGTH_IN_SECONDS:
        num_chunks_per_sample = math.ceil(sample_length / MAX_LENGTH_IN_SECONDS)
        avg_chunk_length = sample_length / num_chunks_per_sample
        num_words = len(batch["words"][0])

        # start chunking by times
        start_word_idx = end_word_idx = 0
        chunk_start_time = 0
        for n in range(num_chunks_per_sample):
            while (end_word_idx < num_words - 1) and (batch["word_start_times"][0][end_word_idx] < segment_start + (n + 1) * avg_chunk_length):
                end_word_idx += 1

            chunk_end_time = int((batch["word_start_times"][0][end_word_idx] - segment_start) * SAMPLE_RATE)
            new_batch["audio"].append(batch["audio"][0][chunk_start_time: chunk_end_time])
            new_batch["text"].append(" ".join(batch["words"][0][start_word_idx: end_word_idx]))

            chunk_start_time = chunk_end_time
            start_word_idx = end_word_idx
    else:
        new_batch["audio"].append(batch["audio"][0])
        new_batch["text"].append(" ".join(batch["words"][0]))

    return new_batch
    
ami = ami.map(chunk_into_max_n_seconds, batched=True, batch_size=1, remove_columns=ami["train"].column_names, num_proc=64)

A segmented and chunked dataset of the config "headset-single"can be found here.

Supported Tasks and Leaderboards

automatic-speech-recognition: The dataset can be used to train a model for Automatic Speech Recognition (ASR). The model is presented with an audio file and asked to transcribe the audio file to written text. The most common evaluation metric is the word error rate (WER). The task does not have an active leaderboard at the moment.
speaker-diarization: The dataset can be used to train model for Speaker Diarization (SD). The model is presented with an audio file and asked to predict which speaker spoke at what time.

Languages

The audio is in English.

Dataset Structure

Data Instances

A typical data point comprises the path to the audio file (or files in the case of the multi-headset or multi-microphone dataset), called file and its transcription as a list of words, called words. Additional information about the speakers, the word_start_time, word_end_time, segment_start_time, segment_end_time is given. In addition

and its transcription, called text. Some additional information about the speaker and the passage which contains the transcription is provided.

{'word_ids': ["ES2004a.D.words1", "ES2004a.D.words2", ...],
 'word_start_times': [0.3700000047683716, 0.949999988079071, ...],
 'word_end_times': [0.949999988079071, 1.5299999713897705, ...],
 'word_speakers': ['A', 'A', ...], 
 'segment_ids': ["ES2004a.sync.1", "ES2004a.sync.2", ...]
 'segment_start_times': [10.944000244140625, 17.618999481201172, ...],
 'segment_end_times': [17.618999481201172, 18.722000122070312, ...],
 'segment_speakers': ['A', 'B', ...], 
 'words', ["hmm", "hmm", ...]
 'channels': [0, 0, ..], 
 'file': "/.cache/huggingface/datasets/downloads/af7e748544004557b35eef8b0522d4fb2c71e004b82ba8b7343913a15def465f"
 'audio': {'path': "/.cache/huggingface/datasets/downloads/af7e748544004557b35eef8b0522d4fb2c71e004b82ba8b7343913a15def465f",
      	   'array': array([-0.00048828, -0.00018311, -0.00137329, ...,  0.00079346, 0.00091553,  0.00085449], dtype=float32),
      	   'sampling_rate': 16000},
}

Data Fields

word_ids: a list of the ids of the words
word_start_times: a list of the start times of when the words were spoken in seconds
word_end_times: a list of the end times of when the words were spoken in seconds
word_speakers: a list of speakers one for each word
segment_ids: a list of the ids of the segments
segment_start_times: a list of the start times of when the segments start
segment_end_times: a list of the start times of when the segments ends
segment_speakers: a list of speakers one for each segment
words: a list of all the spoken words
channels: a list of all channels that were used for each word
file: a path to the audio file
audio: A dictionary containing the path to the downloaded audio file, the decoded audio array, and the sampling rate. Note that when accessing the audio column: dataset[0]["audio"] the audio file is automatically decoded and resampled to dataset.features["audio"].sampling_rate. Decoding and resampling of a large number of audio files might take a significant amount of time. Thus it is important to first query the sample index before the "audio" column, i.e. dataset[0]["audio"] should always be preferred over dataset["audio"][0].

Data Splits

The dataset consists of several configurations, each one having train/validation/test splits:

headset-single: Close talking audio of single headset. This configuration only includes audio belonging to the headset of the person currently speaking.
headset-multi (4 channels): Close talking audio of four individual headset. This configuration includes audio belonging to four individual headsets. For each annotation there are 4 audio files 0, 1, 2, 3.
microphone-single: Far field audio of single microphone. This configuration only includes audio belonging the first microphone, i.e. 1-1, of the microphone array.
microphone-multi (8 channels): Far field audio of microphone array. This configuration includes audio of the first microphone array 1-1, 1-2, ..., 1-8.

In general, headset-single and headset-multi include significantly less noise than microphone-single and microphone-multi.

	Train	Valid	Test
headset-single	136 (80h)	18 (9h)	16 (9h)
headset-multi (4 channels)	136 (320h)	18 (36h)	16 (36h)
microphone-single	136 (80h)	18 (9h)	16 (9h)
microphone-multi (8 channels)	136 (640h)	18 (72h)	16 (72h)

Note that each sample contains between 10 and 60 minutes of audio data which makes it impractical for direct transcription. One should make use of the segment and word start times and end times to chunk the samples into smaller samples of manageable size.

Dataset Creation

All information about the dataset creation can be found here

Curation Rationale

[Needs More Information]

Source Data

Initial Data Collection and Normalization

[Needs More Information]

Who are the source language producers?

[Needs More Information]

Annotations

Annotation process

[Needs More Information]

Who are the annotators?

[Needs More Information]

Personal and Sensitive Information

The dataset consists of people who have donated their voice online. You agree to not attempt to determine the identity of speakers in this dataset.

Datasets:

legacy-datasets
/

ami

Dataset Card for AMI Corpus

Table of Contents

Dataset Description

Dataset Summary

Dataset Preprocessing

Supported Tasks and Leaderboards

Languages

Dataset Structure

Data Instances

Data Fields

Data Splits

Dataset Creation

Curation Rationale

Source Data

Initial Data Collection and Normalization

Who are the source language producers?

Annotations

Annotation process

Who are the annotators?

Personal and Sensitive Information

Considerations for Using the Data

Social Impact of Dataset

Discussion of Biases

Other Known Limitations

Additional Information

Dataset Curators

Licensing Information

Citation Information

TODO

Contributions

TODO