Processing long audio or video files often bogs down a laptop, but Parakeet has introduced an engine under 200MB that can run on consumer CPUs while delivering over 100x speed increases. As part of the "Build with AI" series, a local transcriber project is showcased, capable of converting meeting recordings and podcasts directly into text files.

The focus is on Nvidia Parakeet EDocks, a new speech recognition system that significantly reduces audio processing overhead, delivering full transcription capabilities within a compact file. The system utilizes a compressed mathematical structure at 1.58 bits and supports over two dozen languages. A key advantage is the 113x speed increase when running exclusively on Central Processing Units (CPUs). This allows a 60-minute recording to be processed in approximately half a minute, drastically cutting down wait times in daily workflows.

Developers have optimized the model by reducing the amount of data to be downloaded, leading to higher results with fewer resources. The secret to its efficiency lies in simplifying the neural network's weights to three core values: -1, 0, or +1. This change eliminates the need for expensive Graphics Processing Units (GPUs) and allows the system to automatically detect speech moments, performing audio segmentation.

Installation and Setup

To create a local transcriber, you need to install the FFmpeg library to convert various file formats into WAV format, which is understandable by the model. For Windows users, the command to install FFmpeg is as follows:

# Example command to install FFmpeg (for Windows)
# (The specific command may vary depending on the distribution and installation method)
winget install ffmpeg

Next, you need to install the libraries for application development:

• Streamlit: for creating the user interface (dashboard).
• Moon Reader: for loading the Parakeet Redux model.
• NumPy: for performing array operations in Python.

The command to install these libraries is:

pip install streamlit moon-reader numpy

Application Development

The application will have the title "Parakeet Redux Local Transcriber". The layout can be configured for desktop (wide) or mobile (centered).

The process of handling an input file involves the following steps:

1. Conversion to WAV: The input file (audio or video) is converted to WAV format using FFmpeg. The execution time for this operation is tracked using Python's time library. The FFmpeg command is executed using the subprocess library.
2. Loading Parakeet Redux Model: The Parakeet Redux model is loaded using the Moon Reader library. On the first run, loading the approximately 200MB model will take some time.
3. Transcription: The loaded model is used to transcribe the audio. The process involves loading the speech recognition model (speech_model.load()) and then transcribing (speech_model.transcribe()).
4. Displaying Results: The time taken for audio extraction, AI transcription time, and total execution time are displayed.
5. Segmentation and Formatting: The model automatically identifies audio segments, providing the start and end times for each segment, along with the transcribed text. This data is formatted and stored in the body_transcript variable, which is then written to an output file in Markdown format. The user can also download this file.

Demonstration of Operation

The application is launched with the command streamlit run <application_file_name>.py. A short video lasting about 1 minute is used for testing.

Engine Operation Principle

1. Input Data: Raw audio is fed into the processing pipeline.
2. Conversion: FFmpeg converts the file into audio format.
3. Silence Removal: The Parakeet model removes silent sections to save computational resources.
4. Compressed Encoding: The compressed engine converts the remaining vocal frequencies.
5. Output Data: Accurate words with timestamps for subtitle synchronization.

During local transcription, memory usage is reduced by 85% due to the simplified data structure. The file size is only 178 MB, eliminating the need for expensive GPUs. The model automatically handles punctuation and numbers. Standard processors can perform calculations natively, but the presence of a GPU can speed up the process. Batch scanning of folders with media files is possible.

Comparison and Performance

Based on testing across 25 languages, the Parakeet Redux model demonstrates a lower error rate compared to the baseline version, with approximately 10% errors at a size of 200MB versus 1.2GB for the original model. For long podcasts, the error rate is also lower, with a 0.2% performance improvement compared to the original Parakeet version.

Regarding CPU performance, the original Parakeet showed a 45x speed increase, whereas Parakeet Redux, thanks to its photonic architecture, provides a 113x increase. Compared to the baseline model, execution speed increases by 2.5x. The download size is reduced from 1.2GB to 178MB, decreasing disk space usage by 85%.

Test Results

When processing a 1-minute video file:

• Audio extraction took 0.4 seconds.
• Audio-to-text conversion took 3 seconds.
• The total pipeline execution time was 3 seconds.

Example transcribed text: "The China has once again returned the global tech markets as Deep Sea officially released their new 4.1 flash model." It is noted that company names, such as "Deep Sea" or "GPT", may be transcribed incorrectly if the model has not been trained on these specific terms. However, common English words are recognized accurately.

Practical Application

The model can be used in various scenarios:

• Meeting Transcriber: Ensures 100% data privacy by processing files locally.
• Audio Recording Database: Converts scattered audio notes into a searchable text database.
• Educational Video Analysis: The built-in silence detector allows for breaking down long videos into logical segments.
• Multilingual Audio Processing: Enables processing of mixed-language audio files directly on the device.
• Voice Control Tools: Developers can create command-line tools with minimal latency for voice interaction with code.
• Load Distribution: Offloading processing tasks to the CPU frees up the GPU for working with text models.

Advantages and Limitations

Advantages:

• Simple one-command installation in the terminal.
• Automatic segmentation of long audio by pauses.
• Background processing, minimizing system memory load and allowing multitasking without performance slowdown.

Limitations:

• Strong background noise can affect model accuracy. Clean recordings are required.
• Need to format audio files to the model's specific sampling rate using FFmpeg.

Parakeet Redux is an ultra-lightweight private speech recognition engine that runs entirely offline on basic CPUs. For a guide on local setup and project code, please refer to the first link in the description and pinned comments, which contain detailed documentation and setup instructions.