Eagle Speaker Recognition Engine

Made in Vancouver, Canada by Picovoice

Eagle is an on-device speaker recognition engine. Eagle is:

• Private; All voice processing runs locally.
• Cross-Platform:
  • Linux (x86_64), macOS (x86_64, arm64), Windows (x86_64)
  • Android and iOS
  • Chrome, Safari, Firefox, and Edge
  • Raspberry Pi (4, 3) and NVIDIA Jetson Nano

Compatibility

• Python 3.5 or higher
• Runs on Linux (x86_64), macOS (x86_64, arm64), Windows (x86_64), Raspberry Pi (4, 3), and NVIDIA Jetson Nano.

Installation

pip3 install pveagle

AccessKey

Eagle requires a valid Picovoice AccessKey at initialization. AccessKey acts as your credentials when using Eagle SDKs. You can get your AccessKey for free. Make sure to keep your AccessKey secret. Signup or Login to Picovoice Console to get your AccessKey.

Usage

Eagle has two distinct steps: Enrollment and Recognition. In the enrollment step, Eagle analyzes a series of utterances from a particular speaker to learn their unique voiceprint. This step produces an EagleProfile object, which can be stored and utilized during inference. During the Recognition step, Eagle compares the incoming frames of audio to the voiceprints of all enrolled speakers in real-time to determine the similarity between them.

Speaker Enrollment

Create an instance of the profiler:

import pveagle

# AccessKey obtained from Picovoice Console (https://console.picovoice.ai/)
access_key = "${ACCESS_KEY}"
eagle_profiler = pveagle.create_profiler(access_key)

EagleProfiler is responsible for processing and enrolling PCM audio data, with the valid audio sample rate determined by eagle_profiler.sample_rate. The audio data must be 16-bit linearly-encoded and single-channel.

When passing samples to eagle_profiler.enroll, the number of samples must be at least eagle_profiler.min_enroll_samples to ensure sufficient data for enrollment. The percentage value obtained from this process indicates the progress of enrollment, while the feedback value can be utilized to determine the status of the enrollment process.

def get_next_enroll_audio_data(num_samples):
    pass


percentage = 0.0
while percentage < 100.0:
    percentage, feedback = eagle_profiler.enroll(get_next_enroll_audio_data(eagle_profiler.min_enroll_samples))
    print(feedback.name)

After the percentage reaches 100%, the enrollment process is considered complete. While it is possible to continue providing additional audio data to the profiler to improve the accuracy of the voiceprint, it is not necessary to do so. Moreover, if the audio data submitted is unsuitable for enrollment, the feedback value will indicate the reason, and the enrollment progress will remain unchanged.

speaker_profile = eagle_profiler.export()

The eagle_profiler.export() function produces an EagleProfile object, which can be converted into a binary form using the EagleProfile.to_bytes() method. This binary representation can be saved and subsequently retrieved using the EagleProfile.from_bytes() method.

To reset the profiler and enroll a new speaker, the eagle_profiler.reset() method can be used. This method clears all previously stored data, making it possible to start a new enrollment session with a different speaker.

Finally, when done be sure to explicitly release the resources:

eagle_profiler.delete()

Speaker Recognition

Create an instance of the engine with one or more speaker profiles from the EagleProfiler:

eagle = pveagle.create_recognizer(access_key, speaker_profile)

When initialized, eagle.sample_rate specifies the valid sample rate for Eagle. The expected length of a frame, or the number of audio samples in an input array, is defined by eagle.frame_length.

Like the profiler, Eagle is designed to work with single-channel audio that is encoded using 16-bit linear PCM.

def get_next_audio_frame():
    pass


while True:
    scores = eagle.process(get_next_audio_frame())

The scores array contains floating-point numbers that indicate the similarity between the input audio frame and the enrolled speakers. Each value in the array corresponds to a specific enrolled speaker, maintaining the same order as the speaker profiles provided during initialization. The values in the array range from 0.0 to 1.0, where higher values indicate a stronger degree of similarity.

Finally, when done be sure to explicitly release the resources:

eagle.delete()

Demos

pveagledemo provides command-line utilities for processing real-time audio (i.e. microphone) and files using Eagle.