# Work with YazSes: students, researchers, industry YazSes is a production dictation tool **and** an open research platform for voice-first human-computer interaction. Everything runs on-device, every subsystem sits behind a pluggable seam, and the project is small enough that one person can still understand the whole pipeline — which makes it unusually good material for coursework, theses, papers, and pilots. This page maps what the project offers to each audience. Questions and proposals are welcome in [GitHub Discussions](https://github.com/MSKazemi/yazses/discussions) — the **Research Corner** threads are the standing venue for scientific discussion. ```mermaid flowchart TD Q{What do you want
to walk away with?} --> A[A graded project
or a thesis] Q --> B[A publishable
measurement] Q --> C[A deployment
in my organisation] A --> A1["Pick from the eight scoped projects below
— each has an open issue and an evaluation"] B --> B1["Take an open research question
from the survey pages"] C --> C1["Read the privacy statement and
Meeting Mode, then open a Discussion"] A1 --> D[Open a Discussion to claim it] B1 --> D C1 --> D D --> E["Fast review · co-authorship where earned ·
your name in the release notes"] ``` ## Why YazSes works as a research platform The architecture separates every scientifically interesting component behind a small interface, so an experiment replaces one box without touching the rest: | Seam | Interface | Swap in… | |---|---|---| | Speech-to-text engine | `SttEngine` (`stt/base.py`) | your model — Whisper variants and NVIDIA Parakeet TDT already ship | | Voice-activity detection | `[meeting] vad_backend`, `audio/vad_calibrated.py` | energy gates, Silero, your VAD | | Speaker diarization | `recimport/diarizer.py` | sherpa-onnx today; pyannote seam open ([#71](https://github.com/MSKazemi/yazses/issues/71)) | | Speaker embeddings | `voiceprint/` | ECAPA today; lighter embedders open ([#70](https://github.com/MSKazemi/yazses/issues/70)) | | Activation source | `HotkeyBackend` protocol | keyboard, EMG squeeze (YESP serial), your sensor | | Gaze targeting | `gaze/` backends | MediaPipe iris today; your estimator | Two properties matter for studies: **everything is offline** (no cloud confound, works in a lab without agreements with a third-party processor), and the opt-in [learning corpus](../configuration.md) is **encrypted on-device**, which is the right starting point for an ethics/IRB conversation rather than an obstacle. ## For students: project-sized problems Each of these is a real, wanted contribution with an open issue, a defined evaluation, and a maintainer who reviews quickly. They are sized between a course project and a bachelor/master thesis: | Project | Open issue | Flavour | |---|---|---| | STT benchmark harness — WER/RTF methodology + community results | [#72](https://github.com/MSKazemi/yazses/issues/72) | empirical / reproducibility | | Vocabulary biasing for transducer STT (Parakeet ignores prompts) | [#73](https://github.com/MSKazemi/yazses/issues/73) | speech ML, open design | | Streaming STT: Moonshine engine + latency study | [#74](https://github.com/MSKazemi/yazses/issues/74) | systems / real-time | | Noise suppression in the dictation hot path (DeepFilterNet) | [#69](https://github.com/MSKazemi/yazses/issues/69) | audio DSP | | Speaker-embedding reliability on sub-second windows | [#70](https://github.com/MSKazemi/yazses/issues/70) | speaker modelling | | Diarization accuracy vs speed trade-off (pyannote vs sherpa) | [#71](https://github.com/MSKazemi/yazses/issues/71) | evaluation | | Gaze-aware target selection for dictation ("Glance-Type") | [research background](eye-control.md) | HCI / multimodal | | Dysfluency-friendly transcription for stuttered speech | [accessibility use case](../use-cases/accessibility-rsi-hands-free.md) | accessibility / NLP | Supervisors: if you want to run one of these as a course assignment or thesis, open a Discussion — scoping a variant for your semester is exactly the conversation we want to have. ## For researchers: evaluation, citation, collaboration - **Cite the project** with the repository's `CITATION.cff` (GitHub's *Cite this repository* button), which also references the companion preprint (arXiv:2607.28878) — the [BibTeX entry is on the research index](index.md#how-to-cite). - **Reproducibility**: pinned Python dependencies, CPU-int8 inference, offline models — a described configuration re-runs bit-for-bit on commodity hardware. The benchmark harness ([#72](https://github.com/MSKazemi/yazses/issues/72)) is the intended home for shared methodology. - **Study directions already grounded in the codebase**: multimodal input (gaze + voice + EMG), personal VAD and the cocktail-party problem on short windows, code-switching dictation ([Polyglot](../v2-features.md)), privacy-preserving on-device personalization (encrypted corpus + local fine-tuning), and dictation UX for people with RSI or dysfluent speech. - The [research section](index.md) surveys the literature behind the post-keyboard input roadmap ([eye control](eye-control.md), [voice control](voice-control.md), [muscle & brain interfaces](muscle-brain-control.md)). ## For industry: private dictation you can actually deploy - **Nothing leaves the machine** — see the [privacy statement](../privacy-statement.md). No per-seat cloud subscription, no audio retention, no data-processing agreement needed for the dictation path. - **Meeting transcription on-premises** — [Meeting Mode](../meeting-notes-offline.md) produces speaker-labelled transcripts and minutes without audio ever reaching a third party. - **Scriptable deployment** — plain TOML config, a CLI for everything, systemd/launchd/SCM lifecycle, and (coming) a container image for the headless transcription pipeline ([#76](https://github.com/MSKazemi/yazses/issues/76)). - Pilots and integration questions: open a Discussion or contact the maintainer (see the repository profile). Honest expectations: this is an open-source project, not a vendor with an SLA — what you get is source access, fast issue turnaround, and a say in the roadmap. ## For everyone: the community rhythm - **Research Corner** (Discussions) — a recurring scientific thread: one paper, question, or experiment per thread, tied where possible to something measurable in YazSes. Bring results, not just links. - **[Contributor lanes](https://github.com/MSKazemi/yazses/issues/22)** — the pinned issue groups every open task by interest: GUI, packaging, speech/ML, voice-control features, docs and no-code testing. - **[Benchmarks](../benchmarks.md)** — add a row for your machine once the harness lands.