Also available on request as a PDF. Phone number and referees are omitted here and supplied with applications.

Research interests

Time-domain speech processing; end-to-end acoustic modelling; audio information forensics; speech representation learning; deep learning for audio and speech signals.

Education

2022 – 2026
M.A., Computational Linguistics
University of Isfahan, Isfahan, Iran
Thesis: Speaker Recognition in Persian Spontaneous Speech using Deep Learning. Defended May 2026, summa cum laude. Advisor: Prof. Homa Asadi; co-advisor: Dr. Esfandiar Taheri.
2019 – 2022
B.A., English Language and Literature
University of Qom, Qom, Iran

Research experience

2025 – present
Independent Audio ML Researcher
Self-directed research in time-domain audio and forensics
  • Raw-waveform speaker recognition on VoxCeleb1 with a Raw RWKV-7 architecture and AAMSoftmax loss: 8.63% EER, minDCF 0.6274.
  • Manuscript on neural audio authenticity, Discrete RVQ Audio Watermarking via Hierarchical Residual Injection, targeting IEEE TASLP.
  • Open-source Rust crates for digital signal processing and systems-level AI infrastructure.
2023 – 2026
Research Assistant, Speech Processing
Computational Linguistics Group, University of Isfahan (Prof. Homa Asadi)
  • Acoustic-phonetic and speaker-identity entanglement in time-domain models (RawNet, SincNet) and Wav2Vec 2.0 representations.
  • Speaker-Aware Attention Pooling: 99.58% closed-set accuracy on continuous Persian speech.
  • Pronunciation modelling, forced-alignment robustness and channel-degradation studies on a manually segmented Persian corpus.
  • Laboratory DSP infrastructure in Rust, including dasp-rs.

Publications and preprints

  1. A. Ghanipour Amirhandeh, H. Asadi. “Persian Pronunciation Variation: Predictability and Consequences for Forced Alignment.” Submitted to ICPhS 2027. Code
  2. A. Ghanipour Amirhandeh. “Spectral Wasserstein Gradient Preconditioning for Non-Convex Frequency Optimization in Raw-Waveform Neural Networks.” Under review, IEEE Transactions on Audio, Speech and Language Processing.
  3. A. Ghanipour Amirhandeh. “Adversarial Disentanglement of Speaker Identity and L2 Phonological Errors via Raw-Waveform Architectures.” Submitted to NLP4CALL 2026.
  4. A. Ghanipour Amirhandeh. “Open Baselines and Controls for Conclusion Extraction on PARHAF (PARTAGES CU3).” Building Commons for Clinical LLMs International Conference, 2026. Code
  5. A. Ghanipour Amirhandeh. “rPCEN: Scale-Equivariant Rapid-Release Per-Channel Energy Normalization via Parallel Adjoint Block Scans.” Preprint, 2026. DOI
  6. A. Ghanipour Amirhandeh. “Monotonic Analytic Wavelet Front-Ends with Dynamic Non-Stationary Adaptation for Raw-Waveform Speaker Recognition.” Preprint, 2026. DOI
  7. A. Ghanipour Amirhandeh. “Self-Supervised Speech Representations for Low-Resource and On-Device Streaming ASR: Architectural Paradigms, Compression Bottlenecks, and Open Challenges.” Technical report, 2026. DOI
  8. H. Asadi, A. Ghanipour Amirhandeh. “Evaluating the Contribution of Speech Rhythm Measures to MFCC-Based Speaker Recognition.” In preparation, targeting Interspeech 2027. Code
  9. A. Ghanipour Amirhandeh. “Discrete RVQ Audio Watermarking via Hierarchical Residual Injection.” Manuscript in preparation, targeting IEEE TASLP.
  10. A. Ghanipour Amirhandeh. “Speaker Recognition in Persian Spontaneous Speech using Deep Learning.” M.A. thesis, University of Isfahan, 2026 · defended May 2026, summa cum laude.

Professional experience

Apr 2026 – present
AI Engineer
Donget Digital Ltd, London, UK — expense-splitting app in 87 languages
  • Designed a cloud computer-vision pipeline as constrained structured extraction, with schema-constrained decoding and spatial grounding for financial data.
  • Developed an on-device, script-agnostic text classifier (hashed character n-grams, int8, ~4 MB, pure Dart) for 87 languages including low-resource ones, running in real time.
Aug 2025 – Apr 2026
Software Engineer
Donget Digital Ltd, London, UK
  • Built the Flutter client (Android, iOS, Linux, macOS, Windows, web), the Supabase backend, and evaluation and regression tooling for all AI features.

Open-source software

Twelve crates on crates.io (~6,000 downloads); 70+ public repositories. Contributions to OpenAI harmony and openai-python and to uutils/coreutils. Speech and NLP libraries in Rust: dasp-rs (1,600+ downloads), rustam, textgrid (800+), crfrs, crftag, nerrs, arc-eager, cranium. Details on the software page.

Technical skills

Programming
Python, Rust, Dart, SQL, Bash
Deep learning
PyTorch; time-domain architectures (RawNet, SincNet, linear RNNs/RWKV); self-supervised speech models (Wav2Vec 2.0); vector quantisation (RVQ); margin-based losses (AAMSoftmax)
Speech & signal
Differentiable DSP, acoustic phonetics, speaker verification evaluation (EER, minDCF), forced alignment, continuous speech tokenisation
Engineering
Linux (Arch), Git, Model Context Protocol, LSP integration, deep-learning pipeline design; React, Next.js, Django, Flutter, LaTeX

Service

2025 – present
Data Scientist (volunteer), DataKind — ML and analysis for social-impact partner organisations
2024
Data Scientist (volunteer), CrowdDoing, January – May