OpenVINO Running on LattePanda 3 Delta Single Board Computer (5) - Audio Processing



Audio processing is an essential branch in the field of artificial intelligence, enabling various applications such as speaker recognition, noise reduction, and speech-to-text conversion. This test report will focus on introducing the audio processing projects available in the OpenVINO platform and their application scenarios. OpenVINO provides powerful tools and resources for audio processing tasks, including deep learning models tailored for audio data and efficient inference engines. These optimized models and engines enable real-time, high-quality audio processing on different hardware platforms, offering developers a convenient way to deploy and perform inference.


Within the OpenVINO notebooks, there are several audio-related projects. For instance, analyzing the acoustic features of audio signals can help identify different speakers based on their pronunciation. Additionally, noise reduction techniques can filter out unwanted noise and interference from audio, enhancing speech recognition and communication clarity and accuracy. Furthermore, audio processing includes speech-to-text functionality, which converts spoken content into text form, widely applied in speech recognition, intelligent assistants, and automatic subtitle generation, among other scenarios.



This project provides a complete step-by-step guide for performing speech-to-text recognition using OpenVINO. It includes model downloading, conversion, audio processing, model loading and inference, and output decoding. By following these steps, audio files can be converted into their corresponding text representations.


Project Repository:


Testing Steps

1、Import necessary libraries and modules and install required dependencies.


2、Set up variables such as model folder path, download folder path, data folder path, precision, model name, etc.


3、Download and convert public models:

a. Use the omz_downloader tool to download the selected models.

b. Use the omz_converter tool to convert the downloaded PyTorch models into OpenVINO IR format.


4、Audio processing:

a. Load the audio file.

b. Convert the audio file into a Mel spectrogram.

c. Adjust the Mel spectrogram to the format required for the model's input.


5、Load the model:

a. Create an instance of OpenVINO's Core Engine.

b. Read and load the model.

c. Compile the model for inference.


6、Perform inference:

a. Pass the input to the loaded model and run inference.

b. Obtain the model's output.


7、Decode the output:

a. Perform post-processing on the model's output to convert it into a more readable text format.


Test Results

  • The response speed is relatively fast.


  • The recognition results are accurate.




Speaker diarization is the process of partitioning an audio stream containing human speech into homogeneous segments for each speaker's identity. By using and OpenVINO, it is possible to build a speaker diarization pipeline that can perform speaker separation and recognition on audio files.


Project Repository:


Testing Steps

1、Feature extraction: Convert the raw waveform into audio features, such as mel spectrogram.


2、Speech activity detection: Identify parts of the audio containing speech activity and ignore silence and noise.


3、Speaker change detection: Detect speaker change points in the audio.


4、Voice embedding: Encode each segment by creating feature representations.


5、Speaker clustering: Cluster the segments based on their vector representations. Different clustering algorithms can be applied depending on the availability of the clustering count (k) and the embedding process from the previous step.


Test Results

When using the provided example WAV files, the speaker diarization system can accurately separate different speakers. However, the quality of the audio is crucial for good performance. Audio with high levels of background noise may result in poorer separation quality.



FreeVC allows for voice conversion from the source speaker's voice to the target style while maintaining the linguistic content without the need for text annotations.


Project Repository:


Testing Steps

1、Prior Encoder: Contains a WavLM model, a bottleneck extractor, and a normalizing flow. The WavLM model is used for feature extraction from the audio signal.


2、Speaker Encoder: This component is responsible for extracting speaker embeddings from the input audio.


3、Decoder: Performs voice conversion by synthesizing the converted speech based on the provided embeddings.


Test Results

  • Official example audio file conversions can be found below.
  • The custom audio source is transformed as follows:


In this testing, we explored three speech-related projects, including speech-to-text, multi-speaker diarization, and voice conversion using OpenVINO. The summary of the testing is as follows:


Besides the mentioned projects, OpenVINO offers a powerful set of speech tools and models. With the continuous development of deep learning and neural networks, speech processing techniques will become more efficient and accurate. Through optimization and acceleration with OpenVINO, real-time speech processing can be achieved on edge devices, converting speech into valuable information. Moreover, as speech interaction becomes more prevalent in various fields, OpenVINO's support in speech processing projects will provide more possibilities for research and applications in areas such as speech recognition and synthesis, leading to smarter, more efficient, and convenient speech applications.


In addition to the mentioned projects, OpenVINO offers a wealth of other functionalities. You might be interested in exploring the following articles:

OpenVINO Running on LattePanda 3 Delta Single Board Computer (1) - Object Detection 

OpenVINO Running on LattePanda 3 Delta Single Board Computer (2) - Text Recognition

OpenVINO Running on LattePanda 3 Delta Single Board Computer (3) - NLP 

OpenVINO Running on LattePanda 3 Delta Single Board Computer (4) - Pose Estimation