Crate raqeem_core 

raqeem-core — رقيم. Turn an audio file into Arabic text by delegating inference to a transcription endpoint (Cohere’s hosted API, or your own vLLM). This crate never loads model weights: it decodes nothing, runs no model — it POSTs the audio and folds the result. That is what keeps it light and callable from any language over the CLI.

use raqeem_core::{Endpoint, Transcriber};

let endpoint = Endpoint::cohere(std::env::var("COHERE_API_KEY").unwrap(), None);
let transcript = Transcriber::new(endpoint)
    .language("ar")
    .transcribe(std::path::Path::new("voice_note.ogg"))?;
println!("{}", transcript.text);

Structs

Endpoint

Where and how to reach a transcription server.

Transcriber

Transcribes audio by POSTing it to a configured Endpoint.

Transcript

A completed transcription. text is the model’s verbatim output; the folded text_normalized (see normalize_ar) is what downstream parsers consume.

Enums

Error

Everything that can go wrong on the way from an audio file to a transcript.

OutputFormat

Provider

A known transcription backend.

Constants

COHERE_URL

Cohere’s hosted transcription endpoint.

DEFAULT_COHERE_MODEL

Default model id for the Arabic transcriber. Cohere requires a dated model id — undated aliases (cohere-transcribe-arabic) return HTTP 404 “model not found”. Bump this when Cohere ships a newer dated Arabic transcription model.

DEFAULT_LANGUAGE

Default transcription language (ISO-639-1). Arabic-first, by design.

DEFAULT_TIMEOUT_SECS

Default total-request timeout, in seconds. Generous on purpose: it must cover upload + model inference + download, and CPU inference of a voice note can take many seconds. (reqwest’s blocking client otherwise silently defaults to 30s, which truncates a slow transcription mid-inference.)

Functions

normalize_ar

Fold Arabic text to a matching-stable form. Idempotent; ASCII passes through lower-cased, so it is safe to run unconditionally on any transcript.

Type Aliases

Result