Amirhossein Ghanipour

I work across NLP and speech, and I have never really wanted to pick one. On the language side: large language models and how they are adapted, multilingual representation, and clinical and biomedical text. On the speech side: representation learning, prosody, recognition, and spoofing and deepfake detection.

Multimodal work has pulled me in more recently too — speech and text together, vision-language models. It is where a lot of the interesting problems have moved, and the questions I care about get harder when there is more than one modality in play.

Evaluation is one of the parts I genuinely enjoy, and it has stopped being a side concern in this field. Benchmarks saturate, contamination is difficult to rule out, and review after review — including one from the European Commission’s Joint Research Centre — keeps finding that benchmarks often do not measure the construct they claim to. Judge models turn out to be reliable without being valid. With the AI Act coming into force, validity is becoming something you have to demonstrate rather than assume. That shift is what draws me: it turns evaluation from bookkeeping into a research problem.

Lately I have been working through a series of studies in these areas — speech under channel degradation, whether prosodic structure is reproducible at all, pronunciation modelling for forced alignment, baselines and controls for pulling conclusions out of clinical reports. Each one goes out as code with its results, reproducible from a single command, and I add to it as things finish. What is done so far is on the research page.

dasp-rs began during my research assistant work in my master’s and grew into something considerably larger than I set out to build: a complete audio analysis stack in Rust, covering roughly the ground librosa does, in pure Rust with no system dependencies at all. Time–frequency transforms — STFT, constant-Q, variable-Q, reassigned spectrograms — spectral and MIR features, formant estimation and VAD, YIN and pYIN pitch tracking, DTW and Viterbi decoding, NMF, Griffin-Lim reconstruction, loudness and perceptual weighting, harmonic–percussive separation, and a fairly exhaustive set of music and phonetics notation utilities. It gets used by phoneticians, music-information-retrieval people and audio-ML researchers, and it has passed 1,500 downloads on crates.io. The other one I keep returning to is rustam, an NLP toolkit built specifically for Persian — normalisation and informal-to-formal conversion through to POS tagging, NER and dependency parsing.

Deployment matters to me more than it seems to for most people doing research, and that comes from the engineering half of my work. At Donget I built the multilingual model that runs offline and on-device inside the app. The constraints there are not the ones you get in a lab: a fixed memory and storage budget, no GPU, battery and thermal limits, quantisation and the trade-off between model size and accuracy, cold-start latency, and inference that has to stay responsive on two or three cores while the rest of the app is still running. Getting a model to work is one problem. Getting it to work on someone’s phone without flattening the battery is a different one, and it has taught me a fair amount about the first.

I also contribute to open-source AI tooling, with patches merged into openai/harmony and work on the OpenAI Python SDK — race conditions, streaming edge cases, offline tokenizer loading, the sort of thing that only shows up once people are actually running the library. The rest of my time below the research stack goes to Linux: a distribution, kernel modules and driver patches for hardware the kernel does not fully support yet. Mostly Rust, some Go and C, and it is all on the software page.

I am applying for PhD positions in NLP and speech.