Toward Complete Speaker Anonymization: a Streaming Cascaded System for Voice Privacy and PII Removal

1. INTRODUCTION

• The proposed ASR–NLP–TTS cascade combines speaker anonymization with PII detection, an approach whose combination had seen little exploration.Prior systems used ASR+TTS for anonymization [12] or ASR+NER for PII redaction.Source: §1 · p. 1
• Streaming voice privacy may be useful in remote therapy and support hotlines, where identity protection may lower barriers to service use and facilitate more open communication.Source: §1 · p. 1
• The system is designed for almost-streaming use with fast components, as a baseline toward fully anonymized streaming systems.Source: §1 · p. 1
• The paper evaluates the proposed system as part of its contribution.Source: §1 · p. 1

2.1. Automatic Speech Recognition (ASR)

• Whisper is used in a custom VAD-based streaming procedure that processes speech-containing blocks up to a 1.3-second recognition window.Near the window boundary, potentially truncated words are removed and their audio is carried into the next block.Source: §2.1 · p. 2

2.2. PII detection with LLM/NER

• PII detection candidates include two Qwen2.5-Instruct models, two GLiNER-based models, and spaCy’s en_core_web_trf pipeline.Source: §2.2 · p. 2

2.3. Text-to-Speech (TTS)

• VoxCPM 2.0 with NanovLLM-VoxCPM is selected for high-fidelity continuation cloning and native streaming support.The final configuration uses cfg 2.0, 10 inference steps, temperature 0.9, and a 10-step generation cap intended to prevent hallucination or long silences; continuation is restarted after specified silence, phrase, or audio-duration conditions.Source: §2.3 · p. 2

3.1. Datasets

• LibriSpeech supplies development and test data for WER and UTMOS, while SLURP and DialogPII are used to evaluate content protection.The evaluation uses 11,514 SLURP training transcripts, 2,974 SLURP test transcripts, and 2,913 transcripts from the English part of DialogPII; SLURP includes audio suitable for streaming experiments.Source: §3.1 · p. 2

3.2.1. Latency

• Latency measurement uses 100 LibriSpeech train-100 files and reports utterance-onset end-to-end latency and time to first audio, but excludes NLP processing from the ASR+TTS pipeline measurement.The measurements also omit playback temporal drift due to differences between input and generated audio durations.Source: §3.2.1 · p. 2

3.2.2. Utility

• Utility is measured by WER on anonymized LibriSpeech, UAR on anonymized IEMOCAP, and UTMOS for speech quality.Emotion UAR is averaged across five validation folds; the WER evaluation applies no PII removal.Source: §3.2.2 · pp. 2–3

3.2.3. Privacy

• Privacy evaluation measures speech anonymization with a semi-informed ASV system’s EER and PII detection as a text task.PII experiments focus on person, email_address, date and time, place_name, and relation entities; models are evaluated block by block with left context on gold transcripts and ASR predictions using micro entity F1 and micro SLU-F1.Source: §3.2.3 · p. 3

4.1. Latency

• The latency results are reported in Table 1, with ASR latency shaped by the 1.3 s maximum recognition window and generated-audio duration affected by phrase generation and speech-rate mismatch.On average, at least half of the maximum recognition window contributes to ASR latency; the first TTS frames may contain silence.Source: §4.1 · p. 3, Table 1 · p. 3

4.2. Utility

• 29.22% UAR indicates emotion preservation similar to existing non-streaming cascaded systems, while reconstructed LibriSpeech test speech reaches 11.44% WER and UTMOS is 3.21±0.44 versus 3.35±0.28 for original speech.ASR alone reaches 7.83% WER and TTS on ground-truth transcripts reaches 10.38% WER; Table 2 covers LibriSpeech WER and IEMOCAP UAR.Source: §4.2 · p. 3, Table 2 · p. 3

4.3. Privacy

• EER≥44% places the proposed system above the baseline and in the same range as existing cascaded systems, as summarized in Table 3.Source: Table 3 · p. 3, §4.3 · p. 4
• On PII detection, SLU F1 is 69.09 with 66.88 recall, while GLiNER Small v2.5 entity F1 is 61.55 with 59.30 recall; spaCy processes in 4 ms versus 46 ms for Qwen 1.5B.Table 4 presents DialogPII and SLURP results; the model settings were not selected systematically, so the results are not a careful capability comparison.Source: §4.3 · p. 4, Table 4 · p. 4

5. CONCLUSION

• 717 ms median utterance-onset latency and 44.18%–47.66% EER in VPC2024 gender-specific trials place the system above evaluated baselines B2-B5 and near existing offline cascades.The system combines speaker anonymization with PII detection for content protection.Source: §5 · p. 4
• On simulated DialogPII streaming, GLiNER Small achieves 61.22% entity F1 at 6 ms latency, compared with Qwen 1.5B at 56.27% and 46 ms.The results emphasize jointly selecting and coordinating cascade components.Source: §5 · p. 4
• The conclusion notes that dictionary-based filtering could be used for PII, while an LLM could additionally rewrite text to hide information.Source: §5 · p. 4
• Future work targets reduced latency through adaptive buffering and coordinated ASR-TTS scheduling, plus streaming NLP robust to recognition errors and incomplete context.The stated motivation is substantial PII-detection degradation on streaming ASR outputs.Source: §5 · p. 4

• Generative AI tools supported software development, debugging, test generation, implementation analysis, and LaTeX table generation.Source: Generative AI Use Disclosure · p. 4
• The tools also provided editorial and critical feedback on author-written manuscript text, including clarity, presentation, and argumentation.Source: Generative AI Use Disclosure · p. 4
• All numerical results derive from the described experiments, and the authors remain responsible for the methodology, results, and conclusions.AI-generated scripts processed and formatted outputs.Source: Generative AI Use Disclosure · p. 4