Cheetah

Made in Vancouver, Canada by Picovoice

Cheetah is an on-device streaming Speech-to-Text engine. Cheetah is:

• offline and runs locally without an internet connection.
• highly-accurate [1].
• compact and computationally-efficient [1].
• cross-platform. Linux (x86_64), Mac (x86_64), Windows (x86_64), web browsers, Android, iOS, Raspberry Pi, and Beagle Bone are supported. Linux (x86_64) is available for personal and non-commercial use free of charge. Other platforms are only available under the commercial license.
• customizable. Allows adding new words and adapting to different contexts (Available only under the commercial license).

Table of Contents

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

License

This repository is provided for personal & non-commercial use only. Refer to LICENSE for details. If you wish to use Cheetah in a commercial product request access here.

Use Cases

Cheetah is meant to be used for open-domain transcription applications that require live user feedback (incremental transcription results). Open-Domain question answering, voice typing, and meeting note-taking are a few examples.

• If real-time feedback (incremental transcription results) are not needed, you should check out Leopard.
• If you need to understand naturally-spoken (complex) commands within a specific domain you should check out Rhino.
• If you need to recognize a few simple voice commands or activate a device using voice you should check out Porcupine.

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

File-Based Demo

The demo transcribes a set of audio files provided as command line arguments. The demo has been tested using Python 3.6. 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_demo.py --audio_paths PATH_TO_AUDIO_FILE_1,PATH_TO_AUDIO_FILE_2,PATH_TO_AUDIO_FILE_3

Realtime Demo

This Cheetah demo records audio from the microphone and transcribes it in real-time:

python demo/python/cheetah_demo_realtime.py

Note: you need to have a working microphone and it needs to be set as the default audio capture device on your computer for the demo to function correctly.

C Demo Application

File-Based Demo

This demo application accepts a set of WAV files as input and returns their transcripts. Note 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 compatibility or correctness of the input audio files. Set the current working directory to the 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

Realtime Demo

This demo records the input audio from a microphone and transcribes in real-time. Running the command from root of the repository, the demo can be built using gcc:

gcc -I include/ -O3 demo/c/cheetah_demo_realtime.c -ldl -lasound -o cheetah_demo_realtime

The usage can be attained by

./cheetah_demo_realtime

It can be used as follows

./cheetah_demo_realtime \
lib/linux/x86_64/libpv_cheetah.so \
AUDIO_DEVICE_NAME \
lib/common/acoustic_model.pv \
lib/common/language_model.pv \
resources/license/cheetah_eval_linux_public.lic

The AUDIO_DEVICE_NAME parameter for the microphone can be found using

arecord -L

Note: you need to have a working microphone.

Integration

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_path = ... # The file is available under lib/common/acoustic_model.pv
language_model_path = ... # The file is available under lib/common/language_model.pv
license_path = ... # The file is available under resources/license/cheetah_eval_linux_public.lic

handle = Cheetah(library_path, acoustic_model_path, language_model_path, license_path)

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

audio = ... # audio data to be transcribed

num_frames = len(audio) / handle.frame_length

transcript = ''

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

When finished, release the acquired resources.

handle.delete()

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_path = ... // The file is available under lib/common/acoustic_model.pv
const char *language_model_path = ... // The file is available under lib/common/language_model.pv
const char *license_path = ... // The file is available under resources/license/cheetah_eval_linux_public.lic
const int32_t endpoint_duration_sec = ... // endpoint duration in seconds. set to '-1' to disable endpointing

pv_cheetah_t *handle;
const pv_status_t status = pv_cheetah_init(
    acoustic_model_path,
    language_model_path,
    license_path,
    endpoint_duration_sec,
    &handle);
if (status != PV_STATUS_SUCCESS) {
    // error handling logic
}

Now the handle can be used to process incoming audio stream. 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();

char *transcript = ... // buffer for storing transcription.

for (int i = 0; i < num_frames; i++) {
    const int16_t *frame = &audio[i * pv_cheetah_frame_length()];
    
    char *partial_transcript;
    bool is_endpoint; // is updated only if endpoint detection is enabled at construction time.
    const pv_status_t status = pv_cheetah_process(handle, frame, &partial_transcript, &is_endpoint);
    if (status != PV_STATUS_SUCCESS) {
        // error handling logic
    }
    
    strcat(transcript, partial_transcript);
    free(partial_transcript);
}

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

strcat(transcript, final_transcript);
free(final_transcript);

Finally, when done be sure to release resources acquired.

pv_cheetah_delete(handle);

Releases

V1.2.0 — January 14th, 2020

• Improved accuracy
• Runtime optimizations

V1.1.0 — September 2nd, 2019

• Real-time decoding
• Improved accuracy
• Runtime optimizations

V1.0.0 — October 30th, 2018

• Initial release.