cjpais/Handy · error · anyhow::Error

Model is not loaded for transcription.

Error message

Model is not loaded for transcription.

What it means

transcribe() first waits on a condvar until any in-progress model load finishes, then requires the engine slot to be occupied. This error means the engine is None at that point: no model was ever loaded, the last load failed, the model was unloaded (idle timeout / manual unload / model switch), or the engine was dropped after a prior panic. Despite the comment, the current code performs no auto-load — it errors immediately.

Solutions

1. Open settings and confirm a model is downloaded and selected; wait for the model-state-changed "loaded" event before recording
2. Check logs/app notifications for a preceding loading_failed error and fix that first (usually re-download)
3. Increase or disable the model unload timeout if the engine keeps getting evicted between recordings
4. If it follows a panic error, simply retry — the next attempt reloads the model

Example fix

// before: record blind, fail when idle-unload evicted the engine
let text = manager.transcribe(audio)?;

// after: ensure a model is active first (loads synchronously if needed)
if manager.get_current_model().is_none() {
    manager.load_model_blocking(&settings.selected_model, None)?;
}
let text = manager.transcribe(audio)?;

// before triggering a recording: ensure a model is actually resident
if manager.get_current_model().is_none() {
    manager.load_model_blocking(&settings.selected_model, None)?;
}

// Rust analog of a type guard: narrow on manager state before transcribe()
fn model_ready(m: &TranscriptionManager) -> bool {
    m.get_current_model().is_some() && !m.is_loading()
}

match manager.transcribe(audio) {
    Err(e) if e.to_string() == "Model is not loaded for transcription." => {
        manager.load_model_blocking(&settings.selected_model, None)?;
        manager.transcribe(audio) // one retry after load
    }
    other => other,
}

When it happens

Trigger: Global shortcut triggers transcription before the first model finishes downloading/loading; the selected model's load failed (a loading_failed event preceded this); ModelUnloadTimeout elapsed since last use and the engine was unloaded; a previous transcription panicked (error 57) and the engine was not returned.

Common situations: Fresh install with no model downloaded yet; user switched models in settings and recorded during the swap; aggressive unload-timeout setting on a slow machine; earlier load failure user dismissed.

Related errors

• Model failed to load after auto-load attempt. Please check…
• Model not available
• Canary transcription failed
• Cohere transcription failed
• Complete model directory not found

AI-assisted analysis of cjpais/Handy@ba10ce1943 (2026-08-17). Data as JSON: /api/errors/804d59fc741194d4. Report an issue: GitHub.

Appendix: source

Thrown at src-tauri/src/managers/transcription.rs:1208

        debug!("Audio vector length: {}", audio_len);

        if audio.is_empty() {
            debug!("Empty audio vector");
            self.maybe_unload_immediately("empty audio");
            return Ok(String::new());
        }

        // Check if model is loaded, if not try to load it
        {
            // If the model is loading, wait for it to complete.
            let mut is_loading = self.is_loading.lock().unwrap();
            while *is_loading {
                is_loading = self.loading_condvar.wait(is_loading).unwrap();
            }

            let engine_guard = self.lock_engine();
            if engine_guard.is_none() {
                return Err(anyhow::anyhow!("Model is not loaded for transcription."));
            }
        }

        // Get current settings for configuration
        let settings = get_settings(&self.app_handle);

        // Validate selected language against the model's supported languages.
        // If the language isn't supported, fall back to "auto" to prevent errors.
        // Validate against the model that's actually loaded (which can differ
        // from settings.selected_model when a caller loaded a specific model —
        // e.g. the --transcribe-file path's --model), not the persisted
        // selection.
        let active_model = self
            .get_current_model()
            .unwrap_or_else(|| settings.selected_model.clone());
        // Resolve the persisted language *intent* into the language this model
        // will actually use. The coercion is capability-aware (a must-pick model
        // never receives "auto") and computed fresh here — it is never written

View on GitHub (pinned to ba10ce1943)