Press shortcut → speak → get text. Free and open source ❤️

Whispering is an open-source speech-to-text application. Press a keyboard shortcut, speak, and your words will transcribe, transform, then copy and paste at the cursor.

I really like hands-free voice dictation. For years, I relied on transcription tools that were almost good, but they were all closed-source. Even those claiming to be "local" or "on-device" were still black boxes that left me wondering where my audio really went.

So I built Whispering. It's open-source, local-first, and most importantly, transparent with your data. ​​Your data is stored locally on your device, and your audio goes directly from your machine to a local provider (Whisper C++, Speaches, etc.) or your chosen cloud provider (Groq, OpenAI, ElevenLabs, etc.) without any middleman or vendor lock-in. For me, the features were good enough that I left my paid tools behind (I used Superwhisper and Wispr Flow before).

Productivity apps should be open-source and transparent with your data, but they also need to match the UX of paid, closed-software alternatives. I hope Whispering is near that point. I use it for several hours a day, from coding to thinking out loud while carrying pizza boxes back from the office.

Here's an overview:

Click to watch: Complete setup and usage guide (5 min)

Note: We recently added support for Whisper C++, which isn't included in the video. For setup, see the docs below.

and here's how I personally am using it with Claude Code these days.

There are plenty of transcription apps out there, each with their own strengths. Whispering has a few tricks up its sleeve, like a voice-activated mode for truly hands-free operation (no button holding), and customizable AI transformations with any prompt/model. The space is full of great ideas, but I just wanted to add some extra competition from the OSS ecosystem.

Built with Svelte 5 and Tauri, so it's tiny (~22MB) and starts instantly. The codebase is well-documented and designed to be understood and audited. That way, you know where your audio goes, how it's processed, and what data is stored. And finally, the cost savings. When you cut out the middleman, you pay providers directly:

We're hoping that together in the open-source, local-first community, we can build something better than any closed-source alternative. The code is open-source because I believe that fundamental tools shouldn't require trusting a black box. Companies pivot, get acquired, or shut down. But open source is forever. ❤️

Set up Whispering and be ready to transcribe in about two minutes.

brew install --cask epicenter-whispering

wget https://github.com/epicenter-md/epicenter/releases/download/v7.7.2/Whispering_7.7.2_amd64_linux.AppImage
chmod +x Whispering_7.7.2_amd64_linux.AppImage
./Whispering_7.7.2_amd64_linux.AppImage

wget https://github.com/epicenter-md/epicenter/releases/download/v7.7.2/Whispering_7.7.2_amd64_linux.deb
sudo dpkg -i Whispering_7.7.2_amd64_linux.deb

wget https://github.com/epicenter-md/epicenter/releases/download/v7.7.2/Whispering-7.7.2-1.x86_64_linux.rpm
sudo rpm -i Whispering-7.7.2-1.x86_64_linux.rpm

Whispering offers two approaches to transcription. Pick the one that fits your needs:

Take your transcription experience to the next level with these advanced features:

Whispering stores as much data as possible locally on your device, including recordings and text transcriptions. This approach ensures maximum privacy and data security. Here's an overview of how data is handled:

1. Local Storage: Voice recordings and transcriptions are stored in IndexedDB, which is used as blob storage and a place to store all of your data like text and transcriptions.

2. Transcription Service: The only data sent elsewhere is your recording to an external transcription service—if you choose one. You have the following options:

   • External services like OpenAI, Groq, or ElevenLabs (with your own API keys)
   • A local transcription service such as Speaches, which keeps everything on-device

3. Transformation Service (Optional): Whispering includes configurable transformation settings that allow you to pipe transcription output into custom transformation flows. These flows can leverage:

   • External Large Language Models (LLMs) like OpenAI's GPT-4, Anthropic's Claude, Google's Gemini, or Groq's Llama models
   • Hosted LLMs within your custom workflows for advanced text processing
   • Simple find-and-replace operations for basic text modifications

   When using AI-powered transformations, your transcribed text is sent to your chosen LLM provider using your own API key. All transformation configurations, including prompts and step sequences, are stored locally in your settings.

You can change both the transcription and transformation services in the settings to ensure maximum local functionality and privacy.

Most apps are middlemen charging $30/month for API calls that cost pennies. With Whispering, you bring your own API key and pay providers directly. Your audio goes straight from your device to the API. No servers in between, no data collection, no subscriptions.

I built this for myself and use it every day. The code is open source so you can verify exactly what it does.

Svelte 5 + Tauri. The app is tiny (~22MB), starts instantly, and uses minimal resources. The codebase is clean and well-documented if you want to learn or contribute.

Yes, use the Speaches provider for local transcription. No internet, no API keys, completely private.

With Groq (my favorite): $0.02-$0.06/hour. With OpenAI: $0.18-$0.36/hour. Local transcription: free forever. I use it several hours daily and pay about $3/month total.

Your recordings stay on your device in IndexedDB. When you transcribe, audio goes directly to your chosen provider using your API key. No middleman servers. For maximum privacy, use local transcription.

Note: we use anonymized event logging with Aptabase, an open-source, privacy-first analytics service. No personal data is attached to any of these events. You can see exactly what events we log in analytics.ts and where they are logged by searching for rpc.analytics.logEvent in our codebase. You can turn this off in settings at any time.

Yes - set up AI transformations to fix grammar, translate languages, or reformat text. Works with any LLM provider.

Desktop: Mac (Intel & Apple Silicon), Windows, Linux. Web: Any modern browser at whispering.epicenter.so.

Open an issue on GitHub. I actively maintain this and respond quickly.

Whispering showcases the power of modern web development as a comprehensive example application:

• Svelte 5: The UI reactivity library of choice with cutting-edge runes system
• SvelteKit: For routing and static site generation
• Tauri: The desktop app framework for native performance
• WellCrafted: Lightweight type-safe error handling
• Svelte Sonner: Toast notifications for errors
• TanStack Query: Powerful data synchronization
• TanStack Table: Comprehensive data tables
• IndexedDB & Dexie.js: Local data storage
• shadcn-svelte: Beautiful, accessible components
• TailwindCSS: Utility-first CSS framework
• Turborepo: Monorepo management
• Rust: Native desktop features
• Vercel: Hosting platform
• Zapsplat.com: Royalty-free sound effects

• React: UI library
• shadcn/ui: Component library
• Chrome API: Extension APIs

Note: The browser extension is temporarily disabled while we stabilize the desktop app.

Whispering uses a clean three-layer architecture with 97% code sharing between desktop and web versions:

• Service Layer: Platform-agnostic business logic with Result types
• Query Layer: Reactive data management with caching
• UI Layer: Clean Svelte 5 components with minimal logic

The architecture achieves extensive code reuse through build-time platform detection, allowing the same codebase to run natively on desktop (via Tauri) and in the browser.

→ For developers: See the detailed Architecture Deep Dive to understand how everything fits together.

1. Clone the repository: git clone https://github.com/epicenter-md/epicenter.git
2. Change into the project directory: cd epicenter
3. Install the necessary dependencies: bun i

To run the desktop app and website:

cd apps/whispering
bun tauri dev

If you have concerns about the installers or want more control, you can build the executable yourself. This requires more setup, but it ensures that you are running the code you expect. Such is the beauty of open-source software!

cd apps/whispering
bun i
bun tauri build

Find the executable in apps/whispering/target/release

We welcome contributions! Whispering is built with care and attention to clean, maintainable code.

• Follow existing TypeScript and Svelte patterns throughout
• Use Result types from the WellCrafted library for all error handling
• Follow WellCrafted best practices: explicit errors with Result<T, E>, structured TaggedError objects, and comprehensive error context
• Study the existing patterns in these key directories:
  • Services Architecture - Platform-agnostic business logic
  • Query Layer Patterns - RPC pattern and reactive state
  • Constants Organization - Type-safe configuration

→ New to the codebase? Start with the Architecture Deep Dive to understand how everything fits together.

We'd love to expand Whispering's capabilities with more transcription and AI service adapters! Here's how to add a new adapter:

Overview of the adapter system:

1. Transcription services (services/transcription/): Convert audio to text
2. Completion services (services/completion/): Power AI transformations in the transformation pipeline
3. Query layer (query/): Provides reactive state management and runtime dependency injection
4. Settings layer: Stores API keys and user preferences

Adding a new transcription service involves four main steps:

1. Create the service implementation in the appropriate transcription subdirectory:

   • Cloud services: src/lib/services/transcription/cloud/ (OpenAI, Groq, Deepgram, ElevenLabs)
   • Local services: src/lib/services/transcription/local/ (WhisperCpp, Parakeet)
   • Self-hosted services: src/lib/services/transcription/self-hosted/ (Speaches)

   // Example: src/lib/services/transcription/cloud/your-service.ts
   import { WhisperingErr, type WhisperingError } from '$lib/result';
   import type { Settings } from '$lib/settings';
   import { Err, Ok, tryAsync, type Result } from 'wellcrafted/result';
   
   // Define your models directly in the service file
   export const YOUR_SERVICE_MODELS = [
   	{
   		name: 'model-v1',
   		description: 'Description of what makes this model special',
   		cost: '$0.XX/hour',
   	},
   	{
   		name: 'model-v2',
   		description: 'A faster variant with different trade-offs',
   		cost: '$0.YY/hour',
   	},
   ] as const;
   
   export type YourServiceModel = (typeof YOUR_SERVICE_MODELS)[number];
   
   export function createYourServiceTranscriptionService() {
   	return {
   		async transcribe(
   			audioBlob: Blob,
   			options: {
   				prompt: string;
   				temperature: string;
   				outputLanguage: Settings['transcription.outputLanguage'];
   				apiKey: string;
   				modelName: (string & {}) | YourServiceModel['name'];
   				// Add any service-specific options
   			},
   		): Promise<Result<string, WhisperingError>> {
   			// Validate API key
   			if (!options.apiKey) {
   				return WhisperingErr({
   					title: '🔑 API Key Required',
   					description: 'Please enter your YourService API key in settings.',
   					action: {
   						type: 'link',
   						label: 'Add API key',
   						href: '/settings/transcription',
   					},
   				});
   			}
   
   			// Make the API call
   			const { data, error } = await tryAsync({
   				try: () => yourServiceClient.transcribe(audioBlob, options),
   				catch: (error) =>
   					WhisperingErr({
   						title: '❌ Transcription Failed',
   						description: error.message,
   						action: { type: 'more-details', error },
   					}),
   			});
   
   			if (error) return Err(error);
   			return Ok(data.text.trim());
   		},
   	};
   }
   
   export const YourServiceTranscriptionServiceLive =
   	createYourServiceTranscriptionService();

   Don't forget to export your service in src/lib/services/transcription/index.ts:

   import { YourServiceTranscriptionServiceLive } from './cloud/your-service';
   
   export {
   	// ... existing exports
   	YourServiceTranscriptionServiceLive as yourservice,
   };

   And add the API key field to the settings schema in src/lib/settings/settings.ts:

   'apiKeys.yourservice': z.string().default(''),

2. Update the service registry in src/lib/services/transcription/registry.ts:

   // Add import for your service icon (as SVG)
   import yourServiceIcon from '$lib/constants/icons/your-service.svg?raw';
   
   // Add import for your models
   import {
     YOUR_SERVICE_MODELS,
     type YourServiceModel,
   } from './cloud/your-service';
   
   // Add to the TranscriptionModel union type
   type TranscriptionModel =
     | OpenAIModel
     | GroqModel
     | ElevenLabsModel
     | DeepgramModel
     | YourServiceModel;
   
   // Add to TRANSCRIPTION_SERVICE_IDS array
   export const TRANSCRIPTION_SERVICE_IDS = [
     'whispercpp',
     'parakeet',
     'Groq',
     'OpenAI',
     'ElevenLabs',
     'Deepgram',
     'speaches',
     'YourService', // Add your service here
   ] as const;
   
   // Add to TRANSCRIPTION_SERVICES array (in the appropriate section)
   export const TRANSCRIPTION_SERVICES = [
     // ... existing services
     // Add in the cloud services section:
     {
       id: 'YourService',
       name: 'Your Service Name',
       icon: yourServiceIcon,
       invertInDarkMode: true, // or false, depending on your icon
       description: 'Description of what makes your service special',
       models: YOUR_SERVICE_MODELS,
       defaultModel: YOUR_SERVICE_MODELS[0],
       modelSettingKey: 'transcription.yourservice.model',
       apiKeyField: 'apiKeys.yourservice',
       location: 'cloud', // or 'local' or 'self-hosted'
     },
     // ... rest of services
   ] as const satisfies SatisfiedTranscriptionService[];

3. Wire up the query layer in src/lib/query/transcription.ts:

   // Add to the switch statement in transcribeBlob function
   case 'YourService':
     return services.transcriptions.yourservice.transcribe(blob, {
       outputLanguage: settings.value['transcription.outputLanguage'],
       prompt: settings.value['transcription.prompt'],
       temperature: settings.value['transcription.temperature'],
       apiKey: settings.value['apiKeys.yourservice'],
       modelName: settings.value['transcription.yourservice.model'],
     });

4. Update the settings UI in src/routes/(config)/settings/transcription/+page.svelte:

   <!-- Add after other service conditionals -->
   {:else if settings.value['transcription.selectedTranscriptionService'] === 'YourService'}
     <LabeledSelect
       id="yourservice-model"
       label="YourService Model"
       items={YOUR_SERVICE_MODELS.map((model) => ({
         value: model.name,
         label: model.name,
         ...model,
       }))}
       bind:selected={
         () => settings.value['transcription.yourservice.model'],
         (selected) => settings.updateKey('transcription.yourservice.model', selected)
       }
       renderOption={renderModelOption}
     />
     <YourServiceApiKeyInput />
   {/if}

   Create the API key input component in src/lib/components/settings/api-key-inputs/YourServiceApiKeyInput.svelte:

   <script lang="ts">
   	import { LabeledInput } from '$lib/components/labeled/index.js';
   	import { Button } from '$lib/components/ui/button/index.js';
   	import { settings } from '$lib/stores/settings.svelte';
   </script>
   
   <LabeledInput
   	id="yourservice-api-key"
   	label="YourService API Key"
   	type="password"
   	placeholder="Your YourService API Key"
   	value={settings.value['apiKeys.yourservice']}
   	oninput={({ currentTarget: { value } }) => {
   		settings.updateKey('apiKeys.yourservice', value);
   	}}
   >
   	{#snippet description()}
   		<p class="text-muted-foreground text-sm">
   			You can find your YourService API key in your <Link
   				href="https://yourservice.com/api-keys"
   				target="_blank"
   				rel="noopener noreferrer"
   			>
   				YourService dashboard
   			</Link>.
   		</p>
   	{/snippet}
   </LabeledInput>

   And export it from src/lib/components/settings/index.ts:

   export { default as YourServiceApiKeyInput } from './api-key-inputs/YourServiceApiKeyInput.svelte';

AI transformations in Whispering use completion services that can be integrated into transformation workflows. Here's how to add a new AI provider:

1. Create the completion service in src/lib/services/completion/:

   // src/lib/services/completion/your-provider.ts
   import { WhisperingErr, type WhisperingError } from '$lib/result';
   import { Err, Ok, tryAsync, type Result } from 'wellcrafted/result';
   
   export function createYourProviderCompletionService() {
   	return {
   		async complete(options: {
   			apiKey: string;
   			model: string;
   			systemPrompt: string;
   			userPrompt: string;
   			temperature?: number;
   		}): Promise<Result<string, WhisperingError>> {
   			// Validate API key
   			if (!options.apiKey) {
   				return WhisperingErr({
   					title: '🔑 API Key Required',
   					description: 'Please add your YourProvider API key.',
   				});
   			}
   
   			// Make the completion request
   			const { data, error } = await tryAsync({
   				try: () => yourProviderClient.complete(options),
   				catch: (error) =>
   					WhisperingErr({
   						title: '❌ Completion Failed',
   						description: error.message,
   						action: { type: 'more-details', error },
   					}),
   			});
   
   			if (error) return Err(error);
   			return Ok(data.text);
   		},
   	};
   }
   
   export const YourProviderCompletionServiceLive =
   	createYourProviderCompletionService();

2. Register the service in src/lib/services/completion/index.ts:

   import { YourProviderCompletionServiceLive } from './your-provider';
   
   export {
   	// ... existing exports
   	YourProviderCompletionServiceLive as yourprovider,
   };

3. Wire up the transformation handler in src/lib/query/transformer.ts:

   // Add a new case in the handleStep function's prompt_transform switch statement
   case 'YourProvider': {
     const { data: completionResponse, error: completionError } =
       await services.completions.yourprovider.complete({
         apiKey: settings.value['apiKeys.yourprovider'],
         model: step['prompt_transform.inference.provider.YourProvider.model'],
         systemPrompt,
         userPrompt,
       });
   
     if (completionError) {
       return Err(completionError.message);
     }
   
     return Ok(completionResponse);
   }

4. Add API key to settings in src/lib/settings/settings.ts:

   'apiKeys.yourprovider': z.string().default(''),

5. Update transformation types to include your provider models and configuration

Always use the WhisperingErr helper for user-facing errors:

// Good: User-friendly error with action
return WhisperingErr({
	title: '⏱️ Rate Limit Reached',
	description: 'Too many requests. Please try again in a few minutes.',
	action: {
		type: 'link',
		label: 'View rate limits',
		href: 'https://yourservice.com/rate-limits',
	},
});

// Handle different error types
if (error.status === 401) {
	return WhisperingErr({
		title: '🔑 Invalid API Key',
		description: 'Your API key appears to be invalid or expired.',
		action: {
			type: 'link',
			label: 'Update API key',
			href: '/settings/transcription',
		},
	});
}

// Use with tryAsync for automatic error mapping
const { data, error } = await tryAsync({
	try: () => apiClient.makeRequest(),
	catch: (error) =>
		WhisperingErr({
			title: '❌ Request Failed',
			description: error.message,
			action: { type: 'more-details', error },
		}),
});

Create a test file alongside your service:

// Example: src/lib/services/transcription/cloud/your-service.test.ts
import { describe, it, expect } from 'vitest';
import { createYourServiceTranscriptionService } from './your-service';

describe('YourService Transcription', () => {
	it('should handle missing API key', async () => {
		const service = createYourServiceTranscriptionService();
		const result = await service.transcribe(new Blob(), {
			apiKey: '',
			// other options
		});

		expect(result.error).toBeDefined();
		expect(result.error?.title).toContain('API Key Required');
	});

	// Add more tests
});

When submitting a PR for a new adapter, include:

• The service implementation with comprehensive error handling
• All type definitions and constants
• Query layer integration
• Settings UI components
• Tests covering success and error cases
• Documentation of any special requirements or limitations
• Example .env entries if needed

We're excited to see what services you'll integrate! Feel free to open an issue first to discuss your adapter idea.

1. Fork the repository
2. Create a feature branch: git checkout -b feature/your-feature-name
3. Make your changes and commit them
4. Push to your fork: git push origin your-branch-name
5. Create a pull request

When preparing a new release, use our version bumping script to update all necessary files:

# Update version across all project files
bun run bump-version <new-version>

# Example:
bun run bump-version 7.0.1

This script automatically updates:

• Root package.json
• App package.json
• Tauri configuration (tauri.conf.json)
• Cargo manifest (Cargo.toml)

After running the script, follow the displayed instructions to commit, tag, and push the changes.

• UI/UX improvements and accessibility enhancements
• Performance optimizations
• New transcription or transformation service integrations

Feel free to suggest and implement any features that improve usability—I'll do my best to integrate contributions that make Whispering better for everyone.

Whispering's local transcription capabilities are powered by amazing open-source projects:

• transcribe-rs by CJ Pais - A unified Rust library providing abstraction over multiple speech-to-text engines (Whisper and Parakeet). This library enables Whispering to seamlessly support different transcription engines with a consistent API.

• Handy by CJ Pais - The original project that birthed transcribe-rs. Handy is a an awesome cross-platform desktop speech-to-text application, and I personally use it especially when developing Whispering locally.

• whisper.cpp by Georgi Gerganov - High-performance C/C++ port of OpenAI's Whisper model that makes local transcription fast and efficient.

Special thanks to CJ Pais for creating and maintaining essential libraries that make local, privacy-preserving transcription possible in Whispering.

Whispering is released under the MIT License. Fork it, modify it, build on top of it, make it yours.

We're built in public, shaped by user feedback. If you encounter any issues or have suggestions for improvements, please open an issue on the GitHub issues tab. I really appreciate it!

• Community Discord: Discord
• Feature Requests, Issues, and Bug Reports: GitHub Issues

This project is supported by amazing people and organizations:

Transcription should be free, open, and accessible to everyone. Join us in making it so.