Voice Dataset Creation

This repo outlines the steps and scripts necessary to create your own text-to-speech dataset for training a voice model. The final output is in LJSpeech format.

Table of Contents

• Create Your Own Voice Recordings
• Create a Synthetic Voice Dataset
• Create Transcriptions for Existing Voice Recordings
• Other Utilities

---

Create Your Own Voice Recordings

Requirements

• Voice Recording Software
• Omni-directional head-mounted microphone
• Good quality audio card

Create a Text Corpus of Sentences

• Create sentences that will be about 3-10 seconds when spoken
• Use LJSpeech format
  • "|" separated values, wav file id then sentence text
  • 100|this is an example sentence

Speak and Record Sentences

• Speak each sentence as written
• Sample rate should be 22050 or greater

Sentence Lengths

Run scripts/wavdurations2csv.sh to chart out sentence length and verify that you have a good distribution of WAV file lengths.

---

Create a Synthetic Voice Dataset

Requirements

• Google Cloud Platform Compute Engine Instance
  • Cloud API access scopes select Allow full access to all Cloud APIs
• Conda

Installation

Create Conda Environment on GCP Instance

conda create -n tts python=3.7
conda activate tts
pip install google-cloud-texttospeech==2.1.0 tqdm pandas

Create a Text Corpus of Sentences

• Create sentences that will be about 3-10 seconds when spoken
• Use LJSpeech format
  • "|" separated values, wav file id then sentence text
  • 100|this is an example sentence

Generate Synthetic Voice Dataset

• python text_to_wav.py tts_generate

Sentence Lengths

Run scripts/wavdurations2csv.sh to chart out sentence length and verify that you have a good distribution of WAV file lengths.

---

Create Transcriptions for Existing Voice Recordings

Requirements

• Adobe Audition or Audacity
• Google Cloud Platform Compute Engine Instance
  • Cloud API access scopes select Allow full access to all Cloud APIs
• Conda

Installation

Create Conda Environment on GCP Instance

conda create -n stt python=3.7
conda activate stt
pip install google-cloud-speech tqdm pandas

Fill out a Datasheet for the Voice Dataset

• Review Datasheets for Datasets by Gebru et al.: https://arxiv.org/pdf/1803.09010.pdf
• Markdown Datasheet: https://github.com/JRMeyer/markdown-datasheet-for-datasets/blob/master/DATASHEET.md

Mark the Speech

In Adobe Audition, open audio file:

• Select Diagnostics -> Mark Audio
• Select the Mark the Speech preset
• Click Scan
• Click Find Levels
• Click Scan again
• Click Mark All
• Adjust audio and silence signal dB and length until clips are between 3-10 seconds

Or, in Audacity, open audio file:

• Select Analyze->Sound Finder
• Adjust audio and silence signal dB and length until clips are between 3-10 seconds

Adjust Markers or Label Boundaries

In Audition:

• Open Markers Tab
• Adjust markers, removing silence and noise to make clip length between 3 to 10 seconds long

In Audition:

• Adjust label boundaries, removing silence and noise to make clip length between 3 to 10 seconds long

Export Markers/Labels and WAVs

In Audition:

• Select all markers in list
• Select Export Selected Markers to CSV and save as Markers.csv
• Select Preferences -> Media & Disk Cache and Untick Save Peak Files
• Select Export Audio of Selected Range Markers with the following options:
  • Check Use marker names in filenames
  • Update Format to WAV PCM
  • Update Sample Type 22050 Hz Mono, 16-bit
  • Use folder wavs_export

Or, in Audacity:

• Select Export multiple...
  • Format: WAV
  • Options: Signed 16-bit PCM
  • Split files based on Labels
  • Name files using Label/Track Name
  • Use folder wavs_export
• Select Export labels to Label Track.txt

Analyze WAVs with Signal to Noise Ratio Colab

• run colabs/voice_dataset_SNR.ipynb
• Clean or remove noisy files

Create Initial Transcriptions with STT

For Audition, using the exported Markers.csv and wavs folder run:

cd scripts
python wav_to_text.py audition

The script generates a new file, Markers_STT.csv.

For Audacity, using the exported Label Track.txt and wavs folder run:

cd scripts
python wav_to_text.py audacity

The script generates a new file, Label Track STT.csv.

Fine-tune Transcriptions

For Audition:

• Delete all markers
• Select Import Markers from File and select file with STT transcriptions: Markers_STT.csv
• Fine-tune the Description field in Markers to exactly match the words spoken

For Audacity:

• Open Label Track STT.txt in a text editor.
• Fine-tune the Labels field in the text file to exactly match the words spoken

Export Markers (Audition only) and WAVs

For Audition:

• Select all markers in list
• Select Export Selected Markers to CSV and save as Markers.csv
• Select Export Audio of Selected Range Markers with the following options:
  • Check Use marker names in filenames
  • Update Format to WAV PCM
  • Update Sample Type 22050 Hz Mono, 16-bit
  • Use folder wavs_export

For Audacity:

• Select Export multiple...
  • Format: WAV
  • Options: Signed 16-bit PCM
  • Split files based on Labels
  • Name files using Label/Track Name
  • Use folder wavs_export

Convert Markers(Audition) or Labels(Audacity) into LJSpeech format

Using the exported Markers.csv(Audition) or Label Track STT.txt (Audacity) and WAVs in wavs_export, scripts/markersfile_to_metadata.py will create a metadata.csv and folder of WAVs to train your TTS model:

For Audition:

python markersfile_to_metadata.py audition

For Audacity:

python markersfile_to_metadata.py audacity

Sentence Lengths

Run scripts/wavdurations2csv.sh to chart out sentence length and verify that you have a good distribution of WAV file lengths.

---

Other Utilities

Upsample WAV file

ffmpeg: resampy: We tested three methods to upsample WAV files from 16,000 to 22,050 Hz. After reviewing the spectrograms, we selected ffmpeg for upsampling as it includes another 2 KHz of high end information when compared to resampy. scripts/resamplewav.sh

scripts/resamplewav.sh

References

• Mozilla TTS: https://github.com/mozilla/TTS
• Automating alignment, includes segment audio on silence, Google Speech API, and recognition alignment: https://github.com/carpedm20/multi-Speaker-tacotron-tensorflow#2-2-generate-korean-datasets
• Pretraining on large synthetic corpuses and fine tuning on specific ones https://twitter.com/garygarywang
• Datasheets for Datasets https://arxiv.org/abs/1803.09010