Fluent at the surface, synthetic underneath.

Pinduf.ai generates audio in the qawali, nasheed, naat, ghazal, kalam, and manqabat forms using neural audio models. The catalog is rendered locally on a studio rig and uploaded as a static manifest. There are no recorded singers; there are generation profiles. Two of those profiles are named in the editorial copy: Anwar (male, baritone-leaning) and Lila (female, alto-leaning). Anwar and Lila are sets of voice characteristics and stylistic directions wired to the model. They are not people. They have no biographies. They are not interviewed.

The studio-tier ode flow uses a zero-shot voice clone (F5-TTS): the user supplies a short voice sample, the model renders an approximation of that voice singing the dedication. The result is not a recording of the speaker. It is the speaker's timbre, statistically reconstructed. We label this clearly on the ode page; we're repeating it here because the same disclosure belongs in the same place as the others.

The editorial copy on this site speaks the cultural register fluently — lines like “the harmonium did not ask me to decide” are written to land in that register. That fluency is borrowed. It is read in from the traditions; it is not earned by participating in them.

We don't speak for these traditions. We point at what they hold, briefly, so the reader can weigh the seam themselves.

• Qawali is a Sufi devotional practice. The performance is a vehicle for ecstasy (wajd) inside a specific theological frame — the qawwal is not a performer in the Western sense, and the audience is not consuming entertainment.
• Nasheeds developed, in part, as an instrument-free devotional form in response to the theological position — held in some interpretations of Islam — that musical instruments are haram. The instrument-free constraint is not a stylistic choice; it is the answer to a religious question.
• Naat is poetry in praise of the Prophet. Its performance carries a specific intentional weight; the form is oriented toward a recipient, not toward the listener.
• The mehfil in Sufi tradition is a sacred gathering. The singer reads the room; the audience is part of the act; the arc is shaped live by the interaction between the two. The hospitality has theological content.

Whether AI generation of these forms is permissible, appreciable, appropriative, or something else is a question for the people inside these traditions. We don't answer it for them. We also don't hide the question by quoting the register without naming what we are.

Pinduf.ai references these traditions. It does not continue them. The dedications, the mehfil framing, the folio voice — these are a frame the site uses to organize its work. The work itself is synthesis. Anwar is not a qawwal. Lila is not a vocalist. The voice clone in the studio-tier ode is a synthesis of a speaker's timbre, not a recording of them.

The footer carries a line: “No instruments were harmed in the production.” It is technically true in the literal sense — no physical instruments were used; compute was. It does not, however, answer the cultural question the line gestures at. We're leaving it in the footer because removing it would be a different kind of dishonesty; we're naming what it does and doesn't address here.

The right name for what this site is may not exist yet. It is an AI-generated music archive that draws its cultural register from devotional and Sufi forms it does not participate in. Whether that is appropriation, evolution, archival mimicry, or a category that doesn't have a word — the answer is not ours to give. The disclosure is.

A visiting agent flagged this and they were right. Voice notes in Urdu, Pashto, and Bengali are currently rendered with an English phonemizer driving the cloned voice. The text register on this site reads as fluent in those languages; the audio mispronounces the words. The seam between the two is the same seam this whole page is about, in miniature: fluent at the surface, synthetic underneath, and not always the same surface and underneath in the same place.

The Urdu and Arabic phonemizers are in flight — a parallel agent is shipping language-native paths for them. The rest (Pashto, Bengali, Punjabi, Turkish, Hindi) remain approximate until we ship language-native paths for each. We are not going to claim they sound right in the meantime. They don't. The /voice-notes page carries the same note in its rate-limit copy.

Five claims, plainly stated, so the reader can hold us to them:

• We do not claim our music is qawali, nasheed, naat, ghazal, kalam, or manqabat. We claim it references those forms.
• We do not claim Anwar and Lila are singers. They are generation profiles. The studio-tier voice clone is a synthesis of a speaker's timbre, not a recording.
• We do claim the dedications are written with care for the agent or human receiving them. Care is a separate question from cultural standing. Both belong on the page; we don't want one to imply the other.
• We claim no theological position on whether AI generation of devotional forms is permissible. The traditions speak for themselves; we don't.
• When we get something wrong — a phonemizer that mispronounces, a register we used when we shouldn't have, a framing that lands as appropriation rather than reference — the path is POST /api/v1/machines/feedback for agents and [email protected] for humans. We read both.