Spych

Pronounced: Speech

Python wrapper for easily accessing the DeepSpeech python package via python (without the DeepSpeech CLI)

Documentation for Spych Functions

Spych - https://connor-makowski.github.io/spych/core.html

Spych Wake - https://connor-makowski.github.io/spych/wake.html

Key Features

• Simplified access to pretrained DeepSpeech models for offline and free speech transcription

Setup

Make sure you have Python 3.6.x (or higher) installed on your system. You can download it here.

Installation

1. Install SoX

• On Debian/Ubuntu

  sudo apt install sox
  

• On Mac (via homebrew)

  brew install sox
  

• On windows (Recommend WSL)

2. Install Spych

pip install spych

3. Get DeepSpeech Model and Score files:

curl -LO https://github.com/mozilla/DeepSpeech/releases/download/v0.9.3/deepspeech-0.9.3-models.pbmm
curl -LO https://github.com/mozilla/DeepSpeech/releases/download/v0.9.3/deepspeech-0.9.3-models.scorer

Examples

Transcribe Existing Audio File

• Spych Docs

from spych import spych

spych_obj=spych(model_file='deepspeech-0.9.3-models.pbmm', scorer_file='deepspeech-0.9.3-models.scorer')

# Convert the audio file to text
print('Transcription:')
print(spych_obj.stt(audio_file='test.wav'))

• Note: A .wav file at the same sample rate as your selected DeepSpeech models is processed the fastest

Record and Transcribe

• Spych Docs

from spych import spych

spych_obj=spych(model_file='deepspeech-0.9.3-models.pbmm', scorer_file='deepspeech-0.9.3-models.scorer')

# Record using your default microphone for 3 seconds
print('Recording...')
my_audio_buffer=spych_obj.record(duration=3)
print('Recording Finished')

# Convert the audio buffer to text
print('You said:')
print(spych_obj.stt(my_audio_buffer))

Process a Function After Hearing a Wake Word (Example Wake Word: computer)

• Spych Wake Docs
• Spych Docs

from spych import spych, spych_wake

model_file='deepspeech-0.9.3-models.pbmm'
scorer_file='deepspeech-0.9.3-models.scorer'

spych_object=spych(model_file=model_file, scorer_file=scorer_file)

def my_function():
    print("Listening...")
    audio_buffer=spych_object.record(duration=3)M
    print("You said:",spych_object.stt(audio_buffer=audio_buffer))

listener=spych_wake(model_file=model_file, scorer_file=scorer_file, on_wake_fn=my_function, wake_word="computer")

listener.start()

Modifying the DeepSpeech Model

• Initialized spych objects contain a fully functional DeepSpeech.Model object inside of them
• You can modify this for each spych object any time after initialization
• DeepSpeech.Model options are documented here

Example:

spych_obj=spych(model_file='deepspeech-0.9.3-models.pbmm')

spych_obj.model.enableExternalScorer('deepspeech-0.9.3-models.scorer')
spych_obj.model.addHotWord('activate',10.0)

Rasberry Pi 4 Setup

1. Install system requirements

sudo apt install sox git python3-pip python3-scipy python3-numpy python3-pyaudio libatlas3-base

2. Install python requirements

pip3 install deepspeech
pip3 install spych --no-deps

3. Get the DeepSpeech model and score files (note Pi must user .tflite model file)

curl -LO https://github.com/mozilla/DeepSpeech/releases/download/v0.9.3/deepspeech-0.9.3-models.tflite
curl -LO https://github.com/mozilla/DeepSpeech/releases/download/v0.9.3/deepspeech-0.9.3-models.scorer

4. Use the examples above substituting the original model file name for the .tflite one