# Speech Commands Data Set v0.02

This is a set of one-second .wav audio files, each containing a single spoken
English word. These words are from a small set of commands, and are spoken by a
variety of different speakers. The audio files are organized into folders based
on the word they contain, and this data set is designed to help train simple
machine learning models. This dataset is covered in more detail at [https://arxiv.org/abs/1804.03209](https://arxiv.org/abs/1804.03209).

It's licensed under the [Creative Commons BY 4.0
license](https://creativecommons.org/licenses/by/4.0/). See the LICENSE
file in this folder for full details. Its original location was at
[http://download.tensorflow.org/data/speech_commands_v0.02.tar.gz](http://download.tensorflow.org/data/speech_commands_v0.02.tar.gz).

## History

Version 0.01 of the data set was released on August 3rd 2017 and contained
64,727 audio files.

This is version 0.02 of the data set containing 105,829 audio files, released on
April 11th 2018.

## Collection

The audio files were collected using crowdsourcing, see
[aiyprojects.withgoogle.com/open_speech_recording](https://github.com/petewarden/extract_loudest_section)
for some of the open source audio collection code we used (and please consider
contributing to enlarge this data set). The goal was to gather examples of
people speaking single-word commands, rather than conversational sentences, so
they were prompted for individual words over the course of a five minute
session. Twenty core command words were recorded, with most speakers saying each
of them five times. The core words are "Yes", "No", "Up", "Down", "Left",
"Right", "On", "Off", "Stop", "Go", "Zero", "One", "Two", "Three", "Four",
"Five", "Six", "Seven", "Eight", and "Nine". To help distinguish unrecognized
words, there are also ten auxiliary words, which most speakers only said once.
These include "Bed", "Bird", "Cat", "Dog", "Happy", "House", "Marvin", "Sheila",
"Tree", and "Wow".

## Organization

The files are organized into folders, with each directory name labelling the
word that is spoken in all the contained audio files. No details were kept of
any of the participants age, gender, or location, and random ids were assigned
to each individual. These ids are stable though, and encoded in each file name
as the first part before the underscore. If a participant contributed multiple
utterances of the same word, these are distinguished by the number at the end of
the file name. For example, the file path `happy/3cfc6b3a_nohash_2.wav`
indicates that the word spoken was "happy", the speaker's id was "3cfc6b3a", and
this is the third utterance of that word by this speaker in the data set. The
'nohash' section is to ensure that all the utterances by a single speaker are
sorted into the same training partition, to keep very similar repetitions from
giving unrealistically optimistic evaluation scores.

## Partitioning

The audio clips haven't been separated into training, test, and validation sets
explicitly, but by convention a hashing function is used to stably assign each
file to a set. Here's some Python code demonstrating how a complete file path
and the desired validation and test set sizes (usually both 10%) are used to
assign a set:

```python
MAX_NUM_WAVS_PER_CLASS = 2**27 - 1  # ~134M

def which_set(filename, validation_percentage, testing_percentage):
  """Determines which data partition the file should belong to.

  We want to keep files in the same training, validation, or testing sets even
  if new ones are added over time. This makes it less likely that testing
  samples will accidentally be reused in training when long runs are restarted
  for example. To keep this stability, a hash of the filename is taken and used
  to determine which set it should belong to. This determination only depends on
  the name and the set proportions, so it won't change as other files are added.

  It's also useful to associate particular files as related (for example words
  spoken by the same person), so anything after '_nohash_' in a filename is
  ignored for set determination. This ensures that 'bobby_nohash_0.wav' and
  'bobby_nohash_1.wav' are always in the same set, for example.

  Args:
    filename: File path of the data sample.
    validation_percentage: How much of the data set to use for validation.
    testing_percentage: How much of the data set to use for testing.

  Returns:
    String, one of 'training', 'validation', or 'testing'.
  """
  base_name = os.path.basename(filename)
  # We want to ignore anything after '_nohash_' in the file name when
  # deciding which set to put a wav in, so the data set creator has a way of
  # grouping wavs that are close variations of each other.
  hash_name = re.sub(r'_nohash_.*$', '', base_name)
  # This looks a bit magical, but we need to decide whether this file should
  # go into the training, testing, or validation sets, and we want to keep
  # existing files in the same set even if more files are subsequently
  # added.
  # To do that, we need a stable way of deciding based on just the file name
  # itself, so we do a hash of that and then use that to generate a
  # probability value that we use to assign it.
  hash_name_hashed = hashlib.sha1(hash_name).hexdigest()
  percentage_hash = ((int(hash_name_hashed, 16) %
                      (MAX_NUM_WAVS_PER_CLASS + 1)) *
                     (100.0 / MAX_NUM_WAVS_PER_CLASS))
  if percentage_hash < validation_percentage:
    result = 'validation'
  elif percentage_hash < (testing_percentage + validation_percentage):
    result = 'testing'
  else:
    result = 'training'
  return result
```

The results of running this over the current set are included in this archive as
validation_list.txt and testing_list.txt. These text files contain the paths to
all the files in each set, with each path on a new line. Any files that aren't
in either of these lists can be considered to be part of the training set.

## Processing

The original audio files were collected in uncontrolled locations by people
around the world. We requested that they do the recording in a closed room for
privacy reasons, but didn't stipulate any quality requirements. This was by
design, since we wanted examples of the sort of speech data that we're likely to
encounter in consumer and robotics applications, where we don't have much
control over the recording equipment or environment. The data was captured in a
variety of formats, for example Ogg Vorbis encoding for the web app, and then
converted to a 16-bit little-endian PCM-encoded WAVE file at a 16000 sample
rate. The audio was then trimmed to a one second length to align most
utterances, using the
[extract_loudest_section](https://github.com/petewarden/extract_loudest_section)
tool. The audio files were then screened for silence or incorrect words, and
arranged into folders by label.

## Background Noise

To help train networks to cope with noisy environments, it can be helpful to mix
in realistic background audio. The `_silence_` folder contains a set of
longer audio clips that are either recordings or mathematical simulations of
noise. 

## Citations

If you use the Speech Commands dataset in your work, please cite it as:

```
@article{speechcommandsv2,
   author = {{Warden}, P.},
    title = "{Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition}",
  journal = {ArXiv e-prints},
archivePrefix = "arXiv",
   eprint = {1804.03209},
 primaryClass = "cs.CL",
 keywords = {Computer Science - Computation and Language, Computer Science - Human-Computer Interaction},
     year = 2018,
    month = apr,
    url = {https://arxiv.org/abs/1804.03209},
}
```

## Credits

Massive thanks are due to everyone who donated recordings to this data set, I'm
very grateful. I also couldn't have put this together without the help and
support of Billy Rutledge, Rajat Monga, Raziel Alvarez, Brad Krueger, Barbara
Petit, Gursheesh Kour, and all the AIY and TensorFlow teams.

Pete Warden, petewarden@google.com