TL;DR
This paper presents a real-time implementation of multi-band frequency compression (MBFC) for hearing aids, aiming to improve speech perception for individuals with sensorineural hearing loss. By adapting a previously effective but computationally intensive technique to a fixed-point DSP, they achieve significant processing speedups while maintaining perceptual quality, making advanced hearing aid features more feasible.
What Problem Does This Paper Address?
Sensorineural hearing impairment often leads to a widening of auditory filters, which causes increased spectral masking and degrades speech perception. While multi-band frequency compression (MBFC) has shown promise in mitigating these effects by concentrating spectral energy, its computational demands have historically limited its real-time application in hearing aids. Previous MBFC implementations either used fixed-frame processing, which can introduce audible artifacts, or pitch-synchronous processing, which is too slow for real-time use due to the complexity of pitch estimation. This paper tackles the challenge of making MBFC computationally efficient enough for real-time hearing aid implementation without sacrificing perceptual quality.
Methods
The core of the paper is the real-time implementation of MBFC on a 16-bit fixed-point DSP (TMS320C5515). The authors adapt the MBFC technique, which involves segmenting the speech spectrum, compressing spectral samples towards the band center, and then resynthesizing the signal.
Key methodological contributions include:
- Fixed-frame processing with LSEE: To overcome the computational burden of pitch-synchronous processing, they employ a fixed-frame approach. To mitigate the artifacts typically associated with fixed-frame processing (like superimposed pitch), they integrate the Griffin-Lim method, specifically a Least Squares Error (LSEE) estimation approach. This method reconstructs the signal from the modified complex spectrum, avoiding the need for explicit pitch estimation.
- Efficient data handling: The implementation utilizes DMA (Direct Memory Access) for efficient data transfer between the ADC, codec, and DSP. A clever buffering strategy with overlap-add operations (specifically 75% overlap using a modified Hamming window) is employed to manage data flow and minimize processing delays.
- DSP optimization: The entire process is optimized for the TMS320C5515 DSP, leveraging its FFT hardware and fixed-point arithmetic. Complex numbers are represented using 4-byte words (16-bit real, 16-bit imaginary) to streamline calculations.
Datasets
Not discussed in the paper. The evaluation appears to be based on subjective listening tests and objective spectrogram analysis of speech samples.
Key Insights and Ideas
- LSEE as a bridge: The Griffin-Lim method, specifically the LSEE variant, proves to be a crucial innovation. It allows for the benefits of spectral modification without the computational overhead of pitch tracking, effectively bridging the gap between high-quality but slow pitch-synchronous methods and artifact-prone but fast fixed-frame methods.
- Real-time feasibility: The successful implementation on a 16-bit fixed-point DSP demonstrates that advanced signal processing techniques like MBFC can be practically deployed in low-power, embedded hearing aid devices.
- Computational efficiency: The authors achieve real-time operation using only about one-tenth of the DSP's computing capacity, leaving significant room for other essential hearing aid functions (e.g., frequency-selective gain, dynamic range compression).
- Perceptual similarity: The processed speech is perceptually similar to that from pitch-synchronous processing, indicating that the LSEE approach effectively preserves the desired audio characteristics.
Key Results
- Real-time operation achieved: The MBFC algorithm was successfully implemented and operated in real-time on a TMS320C5515 DSP.
- Low computational load: The implementation utilized approximately one-tenth of the DSP's computing capacity.
- Perceptual quality: Speech outputs from the LSEE-based real-time implementation were perceptually similar to those from offline pitch-synchronous processing. This was supported by PESQ-MOS scores between 2.5-3.4 for compression factors of 0.6-1.0.
- Low processing delay: The total processing delay was found to be less than 35 ms for a 26 ms window, which is considered acceptable for hearing aid use, especially when combined with lip-reading.
- Performance with compression factor: Previous work cited in the paper showed that a compression factor of 0.6 yielded a mean increase of 16.5% in recognition scores and a mean decrease of 0.89 s in response time for listeners with moderate sensorineural loss.
Limitations and Future Work
- Limited scope of implementation: The current implementation focuses solely on MBFC. For a complete hearing aid solution, other essential processing steps like frequency-selective gain and multi-band dynamic range compression, with listener-specific settable parameters, need to be integrated.
- Dataset limitations: The evaluation relied on informal listening tests and spectrogram analysis rather than large-scale objective datasets.
- Future work: The authors suggest the need to implement additional hearing aid functionalities alongside MBFC. Further optimization for even lower power consumption and exploration of different compression factors or mappings could also be future directions.
Summary updated