stft implementation and lstm implementation

First review-Dissertation-II
11/06/2025
School of Engineering
Department of Electronics and communication
PIP:6002 Dissertation
THIRD REVIEW

Implementation of STFT-Based Audio Processing with LSTM
Acceleration on Cyclone V FPGA
PIP 6002: THIRD REVIEW
Second review-Dissertation-II
11/06/2025
GUIDED BY,
Dr Muthupandi Gandhi,
ASSOCIATE PROFESSOR
PRESENTED BY,
Muskan.S,
20232ESV0002,
M Tech, Embedded and VLSI System.

11/06/2025
ABSTRACT
1.This project implements Short-Time Fourier Transform (STFT)-based audio processing on a Cyclone V FPGA,
integrating Long Short-Term Memory (LSTM) networks to enhance time-frequency analysis.
2.The design incorporates FIR filtering for signal preprocessing, Verilog-based STFT computation, and LSTM
acceleration, ensuring efficient resource utilization and parallel processing on FPGA.
3.The proposed implementation enables real-time spectral analysis of audio signals with improved accuracy and
computational efficiency, demonstrating the feasibility of FPGA-based deep learning acceleration for signal
processing applications.

11/06/2025
PROBLEM STATEMENT
1.Traditional audio processing techniques struggle with real-time spectral analysis due to high
computational complexity and latency, limiting their effectiveness in embedded systems.
2.Implementing STFT on an FPGA requires efficient resource management and optimization to
handle large-scale signal transformations while maintaining accuracy.
3.There is a need for integrating LSTM acceleration to enhance the time-frequency analysis of
audio signals, enabling improved performance in real-time applications.

INFERENCE
Inference:
1.Efficient Real-Time Processing: The implementation of STFT on a Cyclone V FPGA,
combined with LSTM acceleration, enables real-time spectral analysis of audio signals, making it
suitable for low-latency applications.
2.Optimized Resource Utilization: The integration of FIR filtering and hardware-aware
optimizations ensures efficient FPGA resource usage while maintaining computational accuracy
and speed.
3.Enhanced Signal Analysis: The use of LSTM improves the interpretability and performance of
time-frequency analysis, demonstrating the potential of deep learning-based enhancements for
FPGA-based audio processing.
4.Scalability and Future Scope: The project highlights the feasibility of FPGA-accelerated deep
learning for signal processing, paving the way for future improvements such as adaptive filtering,
reduced power consumption, and real-time AI-driven audio enhancement techniques.

11/06/2025
PROPOSED METHOD
1.Preprocessing & FIR Filtering – The input audio signal is filtered using an FIR filter to remove
noise and enhance relevant frequencies.
2.STFT Computation – The signal is divided into frames, and STFT is applied using FFT to obtain
a time-frequency representation, implemented on a Cyclone V FPGA.
3.LSTM Enhancement – An LSTM network processes STFT outputs to capture temporal
dependencies and improve spectral resolution.
4.FPGA Optimization – Parallel processing and pipeline architecture are used to enhance speed,
reduce latency, and optimize resource utilization.
5.Performance Evaluation – The system is tested for accuracy, efficiency, and real-time
processing, demonstrating FPGA-based deep learning acceleration benefits. 🚀

Step 3:extracting the input waveforms

Step 4:generating the stft spectogram

Step 5: plotting the frequency response of first filter

Step 8:Frequency response of the signals

Implementing on Quartus prime
Steps to add the csv dat to quartus prime and simulate it
1.Convert the data to .mif file
2.Add the .mif file to megawizard pulg in manager as ROM data
3.Generate a hdl file

Results interpreted using the .mif files in matlab

Testing and Simulation
To validate functionality, a Verilog testbench was created to
simulate the system. The testbench applies clock and reset signals,
test inputs, and monitors the output. The following is an example
testbench code for validation:

stft implementation and lstm implementation

Key Formulas
1. Short-Time Fourier Transform (STFT): S(t, f) = ∫ x( )w( - t)e^(-
τ τ
j2 f )d , where w is the window function.
π τ τ
2. FIR Filter Output: y[n] = (b_k * x[n-k]) for k = 0 to N-1, where b_k
Σ
are filter coefficients.

Example Parameters for FIR Filters

Hardware Setup
- Configured Cyclone V FPGA using Quartus Prime.
- Integrated USB Blaster for programming and debugging.
- Verified with ModelSim simulations.

Software Tools
- MATLAB for signal processing simulations.
- Quartus Prime for FPGA configuration.
- ModelSim for waveform and timing analysis.

Conclusion and Future Work
- Successful hardware-software co-design.
- Cyclone V FPGA accelerated system.
- Future Work:
- Multi-channel audio support.
- More complex LSTM models.
- Audio classification integration.

References
1.O'Shaughnessy, D. (2000). Speech Communications: Human and Machine. IEEE Press.
2.Allen, J. B. (1977). Short-term spectral analysis, synthesis, and modification by discrete Fourier transform. IEEE
Transactions on Acoustics, Speech, and Signal Processing, 25(3), 235-238.
3.Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735-1780.
4.Lyons, R. G. (2010). Understanding Digital Signal Processing. Prentice Hall.
5.Smith, S. W. (2003). The Scientist and Engineer’s Guide to Digital Signal Processing. California Technical Publishing.
6.Liu, Y., Wang, Y., & Li, S. (2022). FPGA-based implementation of deep learning for real-time signal processing. IEEE
Access, 10, 5423-5435.
7.Sharma, A., Gupta, P., & Verma, R. (2021). A hardware-efficient approach to STFT computation for embedded audio
applications. Journal of Embedded Systems, 15(2), 120-132.
8.IEEE Standard for Floating-Point Arithmetic (IEEE 754-2019). IEEE Computer Society, 2019.

THANK YOU

stft implementation and lstm implementation

More Related Content

Similar to stft implementation and lstm implementation (20)

Recently uploaded (20)

stft implementation and lstm implementation