What if AI could listen
in the language of home?
Pasht0 explores AI that can understand and respond in Pashto, with voice access, oral histories and language learning shaped by the people who use it.
An unfinished future.
Some knowledge lives in the way a story is told. What would AI look like if it understood the language, the culture and the people behind it?
A short film introducing Pasht0, an exploration of language, oral history and community-governed AI.
Start with the person a system struggles to hear.
A useful assistant should understand the language someone actually speaks: the dialect, the code-switching, the context carried by a phrase. Pasht0 begins with a question about access. If AI works much better in some languages than others, who gets to benefit from it?
Anthropology changes the product brief.
Pashto is spoken across Afghanistan, Pakistan and diaspora communities. Oral histories, proverbs, landay poetry and local expertise carry forms of knowledge that a text-only product can miss. My research explores how speech technology might support those practices while respecting the differences within communities.
A language is not a single worldview, and a recording is not automatic permission to train a model. The concept calls for community review, contributor control and separate permission for providing a service and contributing training data.
Build something useful before asking for data.
The proposed starting points are interpreter assistance, pronunciation and conversation practice, voluntary oral-history capture, and compensated contributions of local expertise. Each needs to provide immediate value. A future speech and translation API could help finance affordable access, but buyer demand and pricing remain hypotheses.
The technical plan
Begin with licensed open-weight models and existing Pashto datasets. Establish speech recognition and translation baselines, then test fine-tuning or continued training. Measure tokenization efficiency before proposing a custom tokenizer. Training a foundation model from scratch is a later possibility dependent on evidence and resources.
Evaluate automatic speech recognition (ASR) word error rate by dialect, age, gender, recording noise and device. Use the same held-out evaluation sets to compare translation adequacy, factuality, cultural relevance, task completion and cost per successful interaction. A better benchmark score matters only if people can complete useful tasks.
Stewardship belongs in the architecture.
The proposed system keeps provenance and granular access permissions, separates service consent from training consent, and supports correction, restriction and withdrawal where feasible. CARE principles and Local Contexts inform the research. The concept rejects inferring tribal identity, political affiliation or precise location from speech.
A plan that can prove the idea wrong
The proposed first 90 days include two dialect evaluations, one useful voice prototype, 15 to 25 buyer interviews and a goal of two to three funded pilots. These are proposed tests, not completed milestones. Lack of model improvement, useful retention, rights-cleared data or funded demand would change or stop the plan.
Pasht0 is currently an anthropological and product research concept. No deployed language model, community partnership, trained dataset or measured learning outcome is claimed.
A few useful answers
What is Pasht0?
Pasht0, also described as Pashto AI, is Adnan Khan’s research and product concept for Pashto-first speech and language infrastructure. It explores interpretation, language learning and voluntary oral-history capture.
Has Pasht0 trained or deployed a model?
No trained model or live deployment is claimed in the supplied research. The plan starts with licensed existing models, dialect-aware evaluation and a useful voice prototype.
Would using the product automatically contribute training data?
The concept requires separate permission for service use and training contributions. It proposes provenance, community review, access restrictions and contributor control.
Sources and scope
Source: Adnan’s supplied “Pashto AI: Why First” v4 research and one-page brief, with the accompanying anthropological research. External links are frameworks and related research communities, not project partners or endorsements.
