TL;DR
This paper introduces an improved method for embedding and recovering hidden audio watermarks in speech. By combining a novel zero-phase inverse filtering technique with sign correlation, the authors significantly boost the robustness and accuracy of watermark detection, even when the watermark is very faint and added without knowledge of the original speech.
What Problem Does This Paper Address?
Embedding hidden information within audio, like watermarks, is crucial for copyright protection, tamper detection, and more. The challenge lies in making these watermarks imperceptible to human listeners while ensuring they can be reliably recovered later. Traditional methods often struggle when the watermark is very weak (low signal-to-noise ratio) or when the embedding process doesn't account for the dynamic nature of speech. This paper tackles the problem of robustly recovering a low-energy, asynchronously embedded watermark (specifically, DTMF tones) from speech, even when the original clean speech is unavailable for comparison (blind detection). The core difficulty is distinguishing the faint watermark signal from the complex and variable speech signal, which acts like noise.
Methods
The paper builds upon a previous method that used sign correlation for watermark detection. Instead of a simple linear correlation, sign correlation replaces the actual signal values with their signs, which can improve robustness. The key innovation here is the introduction of zero-phase inverse filtering applied before the sign correlation.
Here's a breakdown:
- Watermark Embedding: A low-level watermark, composed of DTMF tones, is added directly to the speech signal in the time domain:
y(n) = s(n) + λw(n), wheres(n)is speech,w(n)is the watermark, andλis a scaling factor. This is done without knowledge of the speech's time-varying characteristics. - Zero-Phase Inverse Filtering: This is the novel part. The authors design an adaptive filter that aims to "whiten" the speech signal. It works by estimating the short-time spectrum of the speech-plus-watermark signal (
y(n)) and then creating a filter whose frequency response is the inverse of the speech's magnitude response (Hn(ejw) ≈ 1/|Yn(ejw)|). Crucially, this filter has zero phase shift, meaning it doesn't distort the timing of the watermark. This process effectively reduces the energy of the speech signal, particularly in voiced regions, making the watermark more prominent. - Sign Correlation: After inverse filtering, the signal is processed using sign correlation. This involves taking the sign of the filtered signal and correlating it with the known watermark signal. This non-linear operation is shown to provide a significant processing gain, improving the signal-to-noise ratio at the detector's output.
Datasets
The experiments were conducted using:
- TIMIT database: Utterances from twenty male speakers were used. These were concatenated, with short silence periods inserted between them, and then downsampled to 8 kHz.
- DTMF Tones: A sequence of DTMF tones was used to generate the watermark signal.
The paper does not introduce a new benchmark dataset but rather evaluates its method on a standard speech corpus.
Key Insights and Ideas
- Inverse Filtering as a "Speech Suppressor": The zero-phase inverse filter acts like an adaptive "speech suppressor," effectively removing predictable components of the speech signal. This significantly boosts the watermark-to-speech power ratio, especially in voiced speech segments where speech energy is high.
- Zero Phase is Crucial: The authors emphasize that the zero-phase characteristic of the inverse filter is vital. A non-zero phase shift would distort the watermark's timing, hindering detection.
- Sign Correlation's Robustness: Sign correlation, even without the inverse filter, offers a substantial improvement over linear correlation. This is because it's less sensitive to the exact amplitude of the signal and more focused on the presence or absence of the watermark's "shape."
- Combined Power: The real magic happens when zero-phase inverse filtering and sign correlation are combined. This synergistic approach yields the most significant gains in watermark recovery performance.
Key Results
The paper demonstrates significant improvements in watermark detection performance:
- Sign Correlation Alone: Achieved approximately 26 dB greater processing gain compared to linear correlation.
- Inverse Filtering + Sign Correlation: An additional 5 to 6 dB mean improvement was observed, leading to a total gain of roughly 31-32 dB over linear correlation.
- These improvements were measured in terms of the lowest detectable watermark level (detection threshold) as a function of speech duration. The results show that the enhanced method can detect much fainter watermarks.
Limitations and Future Work
- Finite Word Length: The authors acknowledge that the finite word length of the correlator introduces its own noise, which was accounted for in their simulations.
- Watermark Type: The current work focuses on DTMF tones. Future work could explore the application of this method to other types of watermarks.
- Computational Cost: While not explicitly detailed, adaptive inverse filtering can be computationally intensive. Optimizing this aspect for real-time applications might be a future direction.
- Unnatural Flow: The paper mentions that removing all silence periods from speech utterances would create an unnatural flow between sentences, suggesting that careful handling of speech segmentation is important for realistic scenarios.
Summary updated