Add theo's data and contribution scripts

README.md

@@ -2,22 +2,27 @@
 
 A simple audio/speech dataset consisting of recordings of spoken digits in `wav` files at 8kHz. The recordings are trimmed so that they have near minimal silence at the beginnings and ends.
 
-FSDD is an open dataset, which means it will grow overtime as data is contributed. Thus in order to enable reproducibility and accurate citation in scientific journals the dataset is versioned using `git tags`. 
+FSDD is an open dataset, which means it will grow over time as data is contributed. Thus in order to enable reproducibility and accurate citation in scientific journals the dataset is versioned using `git tags`.
 
 ### Current status
-- 2 speakers
-- 1,000 recordings (50 of each digit)
+- 3 speakers
+- 1,500 recordings (50 of each digit)
 - English pronunciations
 
 ### Organization
 Files are named in the following format:
 `{digitLabel}_{speakerName}_{index}.wav`
 Example: `7_jackson_32.wav`
 
+### Contributions
+Please contribute your homemade recordings. All recordings should be 8kHz `wav ` files and be trimmed to have minimal silence. Don't forget to update `metadata.py` with the speaker meta-data.
+
+To add your data, follow the recording instructions in `acquire_data/say_numbers_prompt.py`
+and then run `split_and_label_numbers.py` to make your files.
+
 ### Metadata
 `metadata.py` contains meta-data regarding the speakers gender and accents. 
 
-
 ### Included utilities
 `trimmer.py`
 Trims silences at beginning and end of an audio file. Splits an audio file into multiple audio files by periods of silence.
@@ -29,17 +34,12 @@ A simple class that provides an easy to use API to access the data.
 Used for creating spectrograms of the audio data. Spectrograms are often a useful pre-processing step.
 
 ### Usage
-The test set officially consists of the first 10% of the recordings. Recordings numbered `0-4` (inclusive) are in the test and `5-49` are in the training set. 
-
-### Contributions
-Please contribute your homemade recordings. All recordings should be 8kHz `wav ` files and be trimmed to have minimal silence. Don't forget to update `metadata.py` with the speaker meta-data.
+The test set officially consists of the first 10% of the recordings. Recordings numbered `0-4` (inclusive) are in the test and `5-49` are in the training set.
 
 ### Made with FSDD
 Did you use FSDD in a paper, project or app? Add it here!
 * [https://github.com/Jakobovski/decoupled-multimodal-learning](https://github.com/Jakobovski/decoupled-multimodal-learning)
 
 
-
-
 ### License
 [Creative Commons Attribution-ShareAlike 4.0 International](https://creativecommons.org/licenses/by-sa/4.0/)

acquire_data/say_numbers_prompt.py

@@ -0,0 +1,67 @@
+import time
+import math
+
+"""
+Prompts you to say numbers. 
+
+Start this, and then hit "Record" in Audacity. 
+http://www.audacityteam.org/download/
+When you start Audacity, look in the bottom-left and set the Project Rate (Hz) to 8000.
+
+It takes about 30 minutes to record a full dataset.
+
+Tips:
+- Turn off your screen saver before you start!
+- Try a short recording session first to make sure everything works OK before doing the full recording.
+
+
+When done, export the audio as one big .wav file and use 'split_and_label_numbers.py' 
+to make the labeled dataset.
+"""
+
+DELAY_BETWEEN_NUMBERS = 3
+REPEATS_PER_NUMBER = 3
+
+
+def wait_until(t):
+    while time.time() < t:
+        time.sleep(0.01)
+
+
+def generate_number_sequence():
+    # We want the numbers jumbled up (helps eliminate any previous-number effects)
+    # This function scrambles the numbers in a deterministic way so that we can remember
+    # what the order was later.
+    # A deterministic shuffle makes labeling easy, makes pausing / resuming the experiment easy, etc.
+    nums = [str(i) for i in range(10) for set_num in range(REPEATS_PER_NUMBER)]
+    for i in range(len(nums)):
+        target = int(round(math.pi * i)) % len(nums)
+        (nums[i], nums[target]) = (nums[target], nums[i])
+    return nums
+
+
+def show_numbers():
+    nums = generate_number_sequence()
+
+    print "Get ready..."
+    time.sleep(1)
+
+    t_start = time.time()
+    for i, num in enumerate(nums):
+        if (float(i)/len(nums) * 100) % 10 == 0:
+            print "\n====", float(i)/len(nums)*100, "% done====\n"
+        else:
+            print ""
+
+        t_next_display = t_start + (i + 1) * DELAY_BETWEEN_NUMBERS
+        t_next_blank = t_next_display + 2.5
+
+        # display a number
+        wait_until(t_next_display)
+        print num
+
+        # blank
+        wait_until(t_next_blank)
+
+if __name__ == '__main__':
+    show_numbers()

acquire_data/split_and_label_numbers.py

@@ -0,0 +1,104 @@
+import os
+from collections import defaultdict
+
+import numpy as np
+from scipy.io.wavfile import read, write
+
+from say_numbers_prompt import generate_number_sequence, DELAY_BETWEEN_NUMBERS
+
+"""
+Splits up the audio data you collected in Audacity.
+
+Adjust the CONSTANTS below and run this file.
+
+Labeled audio will appear in the "recordings" dir.
+"""
+
+YOUR_NAME_HERE = 'theo'
+
+# Where did you save your Audacity-exported wav file?
+PATH_TO_AUDIO_FILE = r'C:\Users\theo\Desktop\spoken_numbers_R_8khz.wav'
+
+# Time (seconds) between the beginning of the file and the first number
+# If your output files end up silent, change this number!
+# It may help to look at the beginning of your recording in Audacity to see the offset.
+START_OFFSET = 1.2
+
+# How long it actually took you to say each number, typically 1.5 seconds
+SECS_PER_NUMBER = 3
+
+LABELS = generate_number_sequence()
+
+
+def split_wav(start_offset, secs_between_numbers, secs_per_number):
+    fname = PATH_TO_AUDIO_FILE
+    rate, sound = read(fname)
+
+    if len(sound.shape) > 1:
+        # Audio probably has L and R channels.
+        # Use the left channel only (mono).
+        sound = sound[:, 0]
+
+    samples_between_numbers = int(rate * secs_between_numbers)
+    offset_idx = int(rate*start_offset)
+
+    counts = defaultdict(lambda: 0)
+
+    for i, label in enumerate(LABELS):
+        label = str(label)
+        start_idx = offset_idx + i * samples_between_numbers
+        stop_idx = start_idx + int(rate * secs_per_number)
+
+        if stop_idx > len(sound):
+            raise('Error: Sound ends before expected number of samples reached for index:' + str(i))
+
+        # trim silence
+        digit_audio = sound[start_idx:stop_idx]
+        digit_audio_trimmed = trim_silence(digit_audio)
+
+        # Build filename
+        outfile = label + "_" + YOUR_NAME_HERE + "_" + str(counts[label]) + ".wav"
+        outfile = 'recordings' + os.sep + outfile
+
+        # Write audio chunk to file
+        print "writing", outfile
+        write(outfile, rate, digit_audio_trimmed)
+        counts[label] += 1
+
+
+
+def trim_silence(audio, n_noise_samples=1000, noise_factor=1.0, mean_filter_size=100):
+    """ Removes the silence at the beginning and end of the passed audio data
+    Fits noise based on the last n_noise_samples samples in the period
+    Finds where the mean-filtered magnitude > noise
+    :param audio: numpy array of audio
+    :return: a trimmed numpy array
+    """
+    start = 0
+    end = len(audio)-1
+
+    mag = abs(audio)
+
+    noise_sample_period = mag[end-n_noise_samples:end]
+    noise_threshold = noise_sample_period.max()*noise_factor
+
+    mag_mean = np.convolve(mag, [1/float(mean_filter_size)]*mean_filter_size, 'same')
+
+    # find onset
+    for idx, point in enumerate(mag_mean):
+        if point > noise_threshold:
+            start = idx
+            break
+
+    # Reverse the array for trimming the end
+    for idx, point in enumerate(mag_mean[::-1]):
+        if point > noise_threshold:
+            end = len(audio) - idx
+            break
+
+    return audio[start:end]
+
+
+if __name__ == '__main__':
+    split_wav(START_OFFSET, DELAY_BETWEEN_NUMBERS, SECS_PER_NUMBER)
+

metadata.py

@@ -10,5 +10,10 @@
     	'gender': 'male',
     	'accent': 'BE/French',
     	'language': 'english'
+    },
+    'theo': {
+    	'gender': 'male',
+    	'accent': 'USA/neutral',
+    	'language': 'english'
     }
 }