# raqeem-core
The library behind [`raqeem`](https://github.com/SufficientDaikon/raqeem) (رقيم) — a
lightweight client for
[`CohereLabs/cohere-transcribe-arabic-07-2026`](https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026),
the most accurate open-source Arabic speech-recognition model (dialects + Arabic/English
code-switching), Apache-2.0.
**Inference is always delegated** to an endpoint you choose — Cohere's hosted API or your
own vLLM. This crate loads no model weights: it POSTs the audio to an OpenAI-compatible
`/audio/transcriptions` endpoint and folds the result. That is what keeps it small, and
what lets the same core drive the CLI, the Python wheel, and any future binding.
```rust
use raqeem_core::{Endpoint, Transcriber};
use std::path::Path;
let endpoint = Endpoint::cohere(std::env::var("COHERE_API_KEY")?, None);
let transcript = Transcriber::new(endpoint)? // fails only if the TLS backend won't init
.language("ar") // the default; shown for clarity
.transcribe(Path::new("voice_note.ogg"))?;
println!("{}", transcript.text); // verbatim, for humans
println!("{}", transcript.text_normalized); // Arabic-folded, for parsing
```
Self-hosted instead — same `Transcriber`, different endpoint:
```rust
let endpoint = Endpoint::openai_compatible(
"http://localhost:8000/v1/audio/transcriptions",
"cohere-transcribe-arabic-07-2026",
None, // no key needed
);
```
`text_normalized` runs the output through `normalize_ar` — alef/hamza folding,
taa-marbuta → haa, tatweel and diacritics stripped, invisible format controls removed
(zero-width joiners, bidi marks, BOM), Arabic-Indic and Persian digits → ASCII (`١٢٫٥`
becomes `12.5`, one number) — so downstream matching gets a stable form. `normalize_ar` is
public and idempotent; you can run it on either side of a comparison.
Two things Cohere's API requires, both handled here: the `model` and `language` form fields
must precede the file part, and the model id must be **dated** (undated aliases 404).
The audio is streamed from the file handle rather than read into memory, so peak memory
doesn't scale with the length of the recording. `Transcriber` is `Send + Sync` and holds a
pooled client — share one across threads rather than building one per file.
MSRV 1.86.
The CLI over this crate is [`raqeem`](https://crates.io/crates/raqeem); the Python wheel is
`pip install raqeem`.
Full docs, the CLI, and the Python bindings:
[github.com/SufficientDaikon/raqeem](https://github.com/SufficientDaikon/raqeem).
Apache-2.0 — same as the model. All accuracy credit belongs to Cohere Labs.