Made in Vancouver, Canada by Picovoice
Cheetah is an on-device streaming Speech-to-Text engine. Cheetah is:
- offline: 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 BeagleBone 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).
- Cheetah
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, contact Picovoice.
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) is not needed, see Leopard.
- If you need to understand naturally-spoken (complex) commands within a specific domain, see Rhino.
- If you need to recognize a fixed set of simple voice commands, or activate a device using voice, see Porcupine.
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.
In order to run, Cheetah requires a valid license file ('.lic' extension). To obtain a time-limited evaluation license file, visit Picovoice Console. To obtain a commercial license, contact Picovoice.
First, make sure you have Python 3 installed along with the necessary packages included in the requirements.txt file:
python --version
Python 3.6.8
pip3 install -r requirements.txt
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 --license_path ${PATH_TO_YOUR_CHEETAH_LICENSE_FILE}
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} --license_path ${PATH_TO_YOUR_CHEETAH_LICENSE_FILE}
This Cheetah demo records audio from the microphone and transcribes it in real-time:
python demo/python/cheetah_demo_realtime.py --license_path ${PATH_TO_YOUR_CHEETAH_LICENSE_FILE}
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.
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 \
${PATH_TO_YOUR_CHEETAH_LICENSE_FILE} \
./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 \
${PATH_TO_YOUR_CHEETAH_LICENSE_FILE} \
${PATH_TO_AUDIO_FILE_1} ${PATH_TO_AUDIO_FILE_2} ${PATH_TO_AUDIO_FILE_3}
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 \
${PATH_TO_YOUR_CHEETAH_LICENSE_FILE}
The AUDIO_DEVICE_NAME
parameter for the microphone can be found using
arecord -L
Note: you need to have a working microphone.
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 .lic file is available from Picovoice Console (https://picovoice.ai/console/)
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()
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 the 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 .lic file is available from Picovoice Console (https://picovoice.ai/console/)
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);
- Improved accuracy
- Runtime optimizations
- Real-time decoding
- Improved accuracy
- Runtime optimizations
- Initial release.