TL;DR
This paper introduces a novel steganographic bandwidth extension (BWE) technique for low-bit-rate AMR speech codecs. By cleverly embedding high-frequency audio information into the pitch delay data of existing narrow-band speech signals, it achieves near wide-band quality without requiring changes to the narrow-band decoder, making it backward compatible.
What Problem Does This Paper Address?
Modern telecommunication systems often rely on narrow-band speech codecs (like AMR) that limit audio quality to 3.4 kHz. While wide-band codecs offer superior audio (up to 7 kHz), they are typically not backward compatible with older narrow-band systems. Bandwidth extension (BWE) aims to bridge this gap by reconstructing high-frequency audio information. Previous BWE methods often generated high-frequency content purely on the receiver side, leading to suboptimal quality, especially in noisy environments. Another approach, steganographic BWE, embeds this high-frequency information directly into the narrow-band bitstream. However, existing steganographic BWE methods struggled to maintain adequate sound quality at the very low bit-rates (below 7 kbps) that are crucial for efficient mobile communication. This paper specifically targets improving steganographic BWE for these low-bit-rate AMR modes.
Methods
The core of this work is an enhanced steganographic BWE method for the AMR codec. The key innovation lies in how high-band audio information is embedded:
- Embedding Location: High-frequency spectral information is embedded into the pitch delay data of the AMR codec. This is a departure from methods that might embed data in other parameters.
- Quantization-Based Embedding: An "extended quantization-based method" is used. This involves quantizing the pitch delay parameter in a way that encodes the high-band information. The paper notes that the sensitivity of human hearing to pitch differences is proportional to the ratio of frequencies, allowing for some flexibility in pitch delay values without significant perceptual degradation.
- Adaptive Embedding: The embedding capacity (how many bits can be hidden) is dynamically adjusted. If the pitch delay value in an odd subframe is large, the embedding bit width is extended. Conversely, if the pitch delay is very small, no data is embedded in the subsequent even subframe to avoid quality loss.
- High-Band Signal Synthesis: At the decoder, the high-band signal is synthesized using a combination of harmonic signals (derived from full-wave rectification of the middle-band signal) and noise signals. The choice between harmonic and noise excitation, and their respective gains, are determined by the embedded information.
- Backward Compatibility: The system is designed so that the bitstream generated by the steganographic encoder can be decoded by a standard legacy AMR decoder to produce a narrow-band signal, and by the steganographic BWE decoder to produce a wide-band signal.
Datasets
The paper describes using a dataset of 550 phonetically balanced sentences spoken by 22 Japanese speakers. These sentences were selected from the Continuous Speech Database for Research (Vol. 1) published by the Acoustical Society of Japan. The speech data was 16-bit quantized and 16-kHz sampled, with durations ranging from 6 to 12 seconds, including silence. This data was used to evaluate the objective quality of the decoded speech signals.
Key Insights and Ideas
- Pitch Delay as a Steganographic Carrier: Utilizing pitch delay data for embedding is an elegant way to leverage existing codec parameters without introducing entirely new data streams. The paper demonstrates that this parameter is sufficiently flexible to carry meaningful high-band information.
- Adaptive Embedding for Quality: The adaptive embedding strategy is crucial. By embedding more data when the pitch delay is less sensitive and less data (or none) when it's more sensitive, the method balances embedding capacity with perceptual quality.
- Low Bit-Rate Focus: The paper's specific focus on low-bit-rate AMR modes (below 7 kbps) addresses a critical gap. Many previous BWE techniques might work well at higher bit-rates but fail to deliver acceptable quality in the most constrained scenarios, which are common in mobile communications.
- Objective vs. Subjective Quality: While the paper relies on objective metrics (PESQ), it acknowledges the need for subjective listening tests to fully validate the perceived quality.
Key Results
- Minimal Narrow-band Degradation: When the steganographically embedded bitstream is decoded by a legacy AMR narrow-band decoder, the quality degradation is reported as less than 0.1 MOS-LQO across all tested bit-rate modes.
- Competitive Wide-band Quality: For AMR bit-rates less than 7 kbps, the sound quality of the steganographic BWE speech signals decoded by the new decoder is comparable to that of the standard AMR-WB codec. This is the paper's primary success metric.
- Improved Performance at Low Bit-Rates: The paper explicitly states that the previous steganographic BWE method [6] was inferior at bit-rates below 7 kbps, where this new method does maintain adequate quality.
Limitations and Future Work
- Noise Conditions: The authors acknowledge that the current BWE algorithm has not been quantitatively tested with speech signals containing background noises (e.g., bubble noises, environmental noises). This is a significant area for future work.
- Subjective Evaluation: While objective quality measurements (PESQ) were performed, the paper highlights the necessity of subjective listening tests to confirm the perceived quality of both the steganographic narrow-band signals and the BWE signals.
- Filter Implementation: The paper mentions that all band-pass filters in the encoder and decoder are implemented using 65-tap FIR filters. It suggests that using IIR filters could potentially reduce computational complexity and algorithmic delay.
Summary updated