securecloud-biz/cheetah

On-device speech-to-text engine powered by deep learning

Cheetah

Made in Vancouver, Canada by Picovoice

Cheetah is an on-device speech-to-text engine. It is

• offline and runs locally without an internet connection. Nothing is sent to cloud to fully protect users' privacy.
• compact and computationally-efficient making it suitable for IoT applications.
• highly-accurate.
• cross-platform. Currently Raspberry Pi, Android, iOS, Linux, Mac, and Windows are supported.
• customizable. Allows adding new words and adapting to different contexts.

Table of Contents

• License
• Performance
• Structure of Repository
• Running Demo Applications
  • C Demo Application
  • Python Demo Application
• Integration
  • C
  • Python
• Releases

License

This repository is provided for non-commercial use only. Please refer to LICENSE for details. The license file in this repository is time-limited. Picovoice assures that the license is valid for at least 30 days at any given time.

If you wish to use Cheetah in a commercial product please send an email to [email protected] with a brief description of your use-case. The following table depicts the feature comparison between the free and commercial version of the engine.

License Type	Free	Commercial
Non-Commercial Use	Yes	Yes
Commercial Use	No	Yes
Supported Platforms	Linux	Linux, Mac, Windows, iOS, Android, Raspberry Pi, and various embedded platforms.
Custom Language Models	No	Yes
Compact Language Models	No	Yes
Support	Community Support	Enterprise Support

Performance

A comparison between accuracy and runtime metrics of Porcupine and two other widely-used libraries, PocketSphinx and Mozilla DeepSpeech, is provided here.

Structure of Repository

Cheetah is shipped as an ANSI C shared library. The binary files for supported platforms are located under lib and header files are at include. Bindings are available at binding to facilitate usage from higher-level languages/platforms. Demo applications are at demo. When possible, use one of the demo applications as a starting point for your own implementation. Finally, resources is a placeholder for data used by various applications within the repository.

Running Demo Applications

Python Demo Application

The demo transcribes a set of audio files provided as command line arguments. The demo has been tested using Python 3.5. Note that the files need to be single-channel, 16KHz, and 16-bit linearly encoded. For more information about audio requirements refer to pv_cheetah.h. The following transcribes the WAV file located in the resource directory.

python demo/python/cheetah_demo.py --audio_paths resources/audio_samples/test.wav

In order to transcribe multiple files concatenate their paths using comma as below.

python demo/python/cheetah.py --audio_paths PATH_TO_AUDIO_FILE_1,PATH_TO_AUDIO_FILE_2,PATH_TO_AUDIO_FILE_3

C Demo Application

This demo application accepts a set of WAV files as input and returns their transcripts. Please not that the demo expects the audio files to be WAV, 16KHz, and 16-bit linearly-encoded. It does not perform any verification to assure the correctness of the input audio files. Running the command file from root of the repository the demo can be built using gcc as below.

gcc -I include/ -O3 demo/c/cheetah_demo.c -ldl -o cheetah_demo

The usage can be attained by

./cheetah_demo

Then it can be used as follows

./cheetah_demo ./lib/linux/x86_64/libpv_cheetah.so ./lib/common/acoustic_model.pv ./lib/common/language_model.pv \
./resources/license/cheetah_eval_linux_public.lic ./resources/audio_samples/test.wav

In order to transcribe multiple files append the absolute path to each additional file to the list of command line arguments as follows

./cheetah_demo ./lib/linux/x86_64/libpv_cheetah.so ./lib/common/acoustic_model.pv ./lib/common/language_model.pv \
./resources/license/cheetah_eval_linux_public.lic PATH_TO_AUDIO_FILE_1 PATH_TO_AUDIO_FILE_2 PATH_TO_AUDIO_FILE_3

Integration

C

Cheetah is implemented in ANSI C and therefore can be directly linked to C applications. pv_cheetah.h header file contains relevant information. An instance of Cheetah object can be constructed as follows.

const char *acoustic_model_file_path = ... // The file is available under lib/common/acoustic_model.pv
const char *language_model_file_path = ... // The file is available under lib/common/language_model.pv
const char *license_file_path = ... // The file is available under resources/license/cheetah_eval_linux_spublic.lic

pv_cheetah_t *handle;
const pv_status_t status = pv_cheetah_init(
    acoustic_model_file_path,
    language_model_file_path,
    license_file_path,
    &handle);
if (status != PV_STATUS_SUCCESS) {
    // error handling logic
}

Now the handle can be used to process incoming audio frames. Cheetah accepts single channel, 16-bit PCM audio. The sample rate can be retrieved using pv_sample_rate(). Finally, Cheetah accepts input audio in consecutive chunks (aka frames) the length of each frame can be retrieved using pv_cheetah_frame_length().

const int16_t *audio = ... // audio data to be transcribed
const int audio_length = ... // number of samples in audio

const int num_frames = audio_length / pv_cheetah_frame_length();

for (int i = 0; i < num_frames; i++) {
    const int16_t *frame = &audio[i * pv_cheetah_frame_length()];
    const pv_status_t status = pv_cheetah_process(handle, frame);
    if (status != PV_STATUS_SUCCESS) {
        // error handling logic
    }
}

char *transcription;
const pv_status_t status = pv_cheetah_transcribe(handle, &transcription)
if (status != PV_STATUS_SUCCESS) {
    // error handling logic
}

Finally, when done be sure to release resources acquired.

    free(transcription);
    pv_cheetah_delete(handle);

Python

cheetah.py provides a Python binding for Cheetah library. Below is a quick demonstration of how to construct an instance of it.

library_path = ... # The file is available under lib/linux/x86_64/libpv_cheetah.so
acoustic_model_file_path = ... # The file is available under lib/common/acoustic_model.pv
language_model_file_path = ... # The file is available under lib/common/language_model.pv
license_file_path = ... # The file is available under resources/license/cheetah_eval_linux_public.lic

handle = Cheetah(library_path, acoustic_model_file_path, language_model_file_path, license_file_path)

When initialized, valid sample rate can be obtained using handle.sample_rate. Expected frame length (number of audio samples in an input array) is handle.frame_length.

audio = ... # audio data to be transcribed

num_frames = len(audio) / handle.frame_length

for i in range(num_frames):
    frame = [i * handle.frame_length:(i + 1) * handle.frame_length]
    handle.process(frame)
    
transcript = handle.transcribe()

When done release the acquired resources.

handle.delete()

Releases

v1.0.0 - October 30th, 2018

• Initial release.