Semantic Communications for Speech Signals

I. INTRODUCTION

• Speech semantic communication was still missing, while prior semantic systems discussed here mainly addressed text and image transmission.The introduction contrasts semantic recovery with systems focused on bit- or symbol-level error metrics.Source: §I · p. 1
• DeepSC-S jointly designs speech and channel coding as two deep neural networks to address source distortion and channel effects.The introduction identifies blockwise systems as potentially suboptimal and conventional linear processing as unable to capture many practical-channel imperfections.Source: §I · p. 2
• SE networks learn and extract essential speech information and assign higher weights to it during training, improving recovery accuracy through attention.The SE-based attention mechanism is part of the proposed DeepSC-S transceiver.Source: §I · p. 2
• A model trained under a fixed fading channel and SNR performs under testing channel conditions without network tuning or retraining.The introduction presents this as robustness to dynamic channel environments.Source: §I · p. 2

A. System Expectation

• The system targets recovery of original speech while learning and extracting speech semantics and mitigating distortion and attenuation from practical wireless channels.Speech-characteristic variation makes this a challenging problem.Source: A. System Expectation · p. 2

B. Transmitter

• The transmitter maps a W-sample speech sequence s through separate neural-network speech and channel encoders to produce symbols x for transmission.The encoder parameters are α and β, respectively; transmitted symbols are normalized to satisfy E ∥x∥2 = 1.Source: B. Transmitter · p. 2, Figure 1 · p. 2
• For a single communication link, the channel takes transmitted symbols x and produces received signal y, with independent Gaussian noise w of variance σ2.The channel coefficients are denoted h.Source: B. Transmitter · p. 2

• The receiver uses a channel decoder to mitigate distortion and attenuation, followed by a speech decoder that recovers speech from learned semantic features.Source: C. Receiver · p. 2
• The channel decoder and speech decoder have parameters χ and δ, respectively.Source: C. Receiver · p. 2
• The transceiver’s objective is to recover speech close to the original, using MSE to measure the difference between input and reconstructed signals.Source: C. Receiver · p. 3

D. Performance Metrics

• SDR measures the L2 error between the original and reconstructed speech; higher SDR indicates better recovery quality.Source: D. Performance Metrics · p. 3
• PESQ assesses listening quality at the receiver and scores speech quality from -0.5 to 4.5.Source: D. Performance Metrics · p. 3

III. PROPOSED SEMANTIC COMMUNICATION SYSTEM FOR SPEECH SIGNALS

• DeepSC-S uses an attention-based 2D CNN for speech coding and a 2D CNN for channel coding.Source: §III · p. 3

A. Model Description

• The input speech samples are framed before an attention-based speech encoder learns semantic features.Source: A. Model Description · p. 3
• A channel encoder converts learned features into U, which is reshaped into symbol sequences X for transmission.Source: A. Model Description · p. 3
• The channel layer maps transmitted symbols X to received symbols Y, with channel coefficients H and Gaussian noise W.Source: A. Model Description · p. 3
• The receiver reshapes Y, decodes channel features, and uses an attention-based speech decoder to reconstruct speech of the original input size.Source: A. Model Description · p. 3

B. Speech Encoder and Decoder

• SE-ResNet modules are sequentially connected to construct the speech encoder and decoder, with squeeze-and-excitation assigning greater attention to essential speech features.Source: B. Speech Encoder and Decoder · p. 3

C. Model Training and Testing

• The model objective is to capture essential speech information and work well across various channels and a wide SNR regime.Source: C. Model Training and Testing · p. 3

1) Training Stage:

• Training continues until the MSE loss stops decreasing, and the number of SE-ResNet modules balances encoder/decoder performance with reasonable training time.Source: 1) Training Stage · p. 4
• The training algorithm initializes transmitter and receiver parameters, uses speech samples with fading and fixed-SNR noise, transmits through the channel, computes MSE, and updates parameters simultaneously by SGD.Source: 1) Training Stage · p. 4
• A trained model from a fixed channel condition is tested directly under various fading channels without network tuning or retraining.Source: 1) Training Stage · p. 4

• The test algorithm evaluates the model over a testing channel set and a wide range of SNR values.Source: 2) Testing Stage · p. 4
• For each channel and SNR, it generates Gaussian noise before testing.Source: 2) Testing Stage · p. 4
• The algorithm outputs recovered speech sample sequences for the tested fading channels and SNR values.Source: 2) Testing Stage · p. 4

• The evaluation compares DeepSC-S against traditional communication systems for speech transmission over telephone systems.Source: §IV · p. 4
• The comparison covers AWGN, Rayleigh, and Rician channels.Source: §IV · p. 4
• The experiments assume accurate channel state information.Source: §IV · p. 4

• The Edinburgh DataShare dataset contains more than 10,000 training WAV files and 800 test WAV files sampled at 16 kHz.Source: A. Speech Dataset and Traditional Model · p. 4
• Speech is down-sampled to 8 kHz for telephone-system comparisons, with sequences fixed at W = 16,384 and formed from 128 frames of length 128.Source: A. Speech Dataset and Traditional Model · p. 4
• The traditional speech source-coding baseline uses 64 Kbps PCM with 256 quantization levels.Source: A. Speech Dataset and Traditional Model · pp. 4–5
• Fig. 4 reports MSE loss across AWGN, Rayleigh, and Rician test channels for models trained under various channels.Source: Figure 4 · p. 5
• The traditional channel-coding baseline uses rate-1/3 turbo codes with 512-symbol blocks and five SOVA decoding iterations, alongside 64-QAM modulation.Source: A. Speech Dataset and Traditional Model · p. 5

1) A Robust Model:

• The speech encoder and decoder each use four SE-ResNet modules, while each channel encoder and decoder uses one 2D CNN module with eight kernels.Source: 1) A Robust Model · p. 5
• Table I presents the parameter settings for DeepSC-S in telephone systems.Source: Table I · p. 5
• Under AWGN testing above about 6 dB, the AWGN-trained model has lower MSE loss than the Rayleigh- and Rician-trained models; under Rician testing, all three models achieve MSE below 9 × 10^-7.Source: 1) A Robust Model · p. 5
• The Rician-trained DeepSC-S is selected as a robust model for coping with varied channel environments.Source: 1) A Robust Model · p. 5
• Training uses a fixed SNR of 8 dB; below 8 dB in testing, the AWGN-trained model has higher MSE than the Rayleigh- and Rician-trained models.Source: 1) A Robust Model · p. 5

2) SDR and PESQ Results:

• The SDR and PESQ tests use a model trained under Rician channels at 8 dB SNR.The model is tested against traditional systems for speech transmission over telephone systems.Source: 2) SDR and PESQ Results · p. 5
• DeepSC-S achieves better SDR than traditional systems across AWGN, Rayleigh, and Rician channels, with especially strong gains at low SNR.Its SDR is higher under Rician than AWGN channels because the model was trained under Rician channels.Source: 2) SDR and PESQ Results · p. 5, Figure 5 · p. 6
• DeepSC-S outperforms traditional approaches on PESQ across fading channels and SNRs, particularly in the low-SNR regime.The comparison is shown for AWGN, Rayleigh, and Rician channels.Source: 2) SDR and PESQ Results · p. 5, Figure 6 · p. 6

V. CONCLUSIONS

• DeepSC-S jointly designs speech and channel coding to extract essential speech information, using squeeze-and-excitation-based attention to improve recovery accuracy.The conclusion describes its model as robust to channel variations and reports that it outperforms traditional systems, especially at low SNR.Source: §V · pp. 5–6