Gör mig skånsk

Public demo: gormigskansk.jardenberg.se

This repo contains the standalone browser app behind Gormigskånsk: a small iPhone-friendly demo that takes spoken input in any language, rewrites it into a more Skånska-flavoured Swedish version, and returns audio in a Skånska target voice.

The repo name stayed dialektlab, but the public-facing product and branding are now Gormigskånsk.

This project is intentionally separate from 2026GPT.

What It Does

• records speech in the browser
• also accepts direct text input in the transcript section
• transcribes the audio with OpenAI
• rewrites the utterance into a Skånska-oriented Swedish version while trying to preserve meaning and length
• renders the result with ElevenLabs using the configured target voice
• plays the audio back in the browser

Current UX

• tap once to start recording
• tap again to stop
• or skip the microphone and type text directly in the transcript section
• first interaction may trigger the browser microphone permission dialog
• recording is placed first in the UI so people can start immediately
• processing feedback is shown with staged status text and a progress bar
• source speech can be auto-detected instead of being locked to Swedish
• localized landing pages are available at /, /en, /es, /pl, and /fr

Localized URLs

• https://gormigskansk.jardenberg.se/ for Swedish
• https://gormigskansk.jardenberg.se/en for English
• https://gormigskansk.jardenberg.se/es for Spanish
• https://gormigskansk.jardenberg.se/pl for Polish
• https://gormigskansk.jardenberg.se/fr for French

The chrome and metadata are localized per route. The voice pipeline stays the same: any supported spoken input is turned into Swedish with a Skånska flavour on output.

Stack

• Node + Express
• plain browser frontend in public/
• OpenAI for transcription and dialect rewrite
• ElevenLabs for the default TTS path
• Cloudflare for the public custom domain, DNS, TLS, and edge delivery in front of the app
• Railway for deployment

Default Model Setup

Current defaults bias for latency while keeping decent rewrite quality:

DIALECTLAB_TRANSCRIBE_MODEL=gpt-4o-mini-transcribe
DIALECTLAB_TEXT_MODEL=gpt-5.4-mini
ELEVENLABS_MODEL_ID=eleven_turbo_v2_5
DIALECTLAB_SPEECH_SPEED=0.95

Notes:

• gpt-4o-mini-transcribe handles speech-to-text
• gpt-5.4-mini is the current default rewrite model
• eleven_turbo_v2_5 is now the default TTS model for the current quality/speed balance
• if you want to compare lower latency, switch ELEVENLABS_MODEL_ID to eleven_flash_v2_5
• if you prefer slightly higher TTS fidelity and can tolerate more delay, switch ELEVENLABS_MODEL_ID to eleven_multilingual_v2

Run Locally

1. Copy .env.example to .env
2. Set at least:

OPENAI_API_KEY=...
ELEVENLABS_API_KEY=...
ELEVENLABS_VOICE_ID=...
ELEVENLABS_JJ_VOICE_ID=...

3. Install dependencies:

npm install

4. Start the app:

npm run dev

5. Open http://localhost:8787

Environment Variables

Commonly useful variables:

OPENAI_API_KEY=...
DIALECTLAB_OPENAI_API_KEY=
ELEVENLABS_API_KEY=...
ELEVENLABS_VOICE_ID=...
ELEVENLABS_JJ_VOICE_ID=...
ELEVENLABS_MODEL_ID=eleven_turbo_v2_5

DIALECTLAB_AUDIO_BACKEND=elevenlabs_tts
DIALECTLAB_INPUT_LANGUAGE=
DIALECTLAB_TRANSCRIBE_MODEL=gpt-4o-mini-transcribe
DIALECTLAB_TEXT_MODEL=gpt-5.4-mini
DIALECTLAB_SPEECH_SPEED=0.95
DIALECTLAB_TARGET_DIALECT=Skanska
DIALECTLAB_IP_MAX=12

Notes:

• DIALECTLAB_OPENAI_API_KEY is optional and only overrides OPENAI_API_KEY
• DIALECTLAB_AUDIO_BACKEND=elevenlabs_tts is optional but makes the intent explicit
• leave DIALECTLAB_INPUT_LANGUAGE empty for auto-detection; only set it if you want to force one source language
• HOST and PORT do not need to be set on Railway
• DIALECTLAB_TTS_MODEL and DIALECTLAB_TTS_VOICE only matter on the OpenAI TTS fallback path

Backend Selection

Two audio rendering paths exist:

• openai_tts
• elevenlabs_tts

Selection rules:

• if DIALECTLAB_AUDIO_BACKEND is set, that wins
• otherwise, if ELEVENLABS_API_KEY exists, the app uses elevenlabs_tts
• otherwise, it falls back to openai_tts

iPhone Notes

Microphone access in mobile Safari generally requires a secure context. That means:

• http://localhost works on the same device
• a deployed HTTPS URL works
• a tunnel also works

Plain local-network http://192.168.x.x:8787 may not allow microphone capture on iPhone.

Railway

The app is already shaped correctly for Railway:

• binds to 0.0.0.0
• respects Railway PORT
• exposes GET /health

Recommended Railway variables:

OPENAI_API_KEY=...
ELEVENLABS_API_KEY=...
ELEVENLABS_VOICE_ID=CuaAIFbkzX2kaNH5EtHZ
ELEVENLABS_JJ_VOICE_ID=Z2ic84BAXoRvhU0mvJIa
ELEVENLABS_MODEL_ID=eleven_turbo_v2_5
DIALECTLAB_AUDIO_BACKEND=elevenlabs_tts
DIALECTLAB_TRANSCRIBE_MODEL=gpt-4o-mini-transcribe
DIALECTLAB_TEXT_MODEL=gpt-5.4-mini
DIALECTLAB_SPEECH_SPEED=0.95
DIALECTLAB_TARGET_DIALECT=Skanska
DIALECTLAB_IP_MAX=12

Production Edge

The public deployment is not just Railway. In production, gormigskansk.jardenberg.se is exposed through Cloudflare in front of the Railway service. In practice that means:

• Cloudflare handles the public custom domain, DNS, TLS termination, and edge-facing delivery
• Railway runs the Node app itself
• the app still controls its own HTML freshness with Cache-Control: no-store on locale-rendered pages

Branding Assets

• favicon: public/favicon.svg
• social preview source: public/og-card.svg
• social preview PNG: public/og-card.png

Page metadata, canonical URLs, and hreflang tags are rendered from lib/locales.js and lib/renderPage.js.

Voice Presets

When ElevenLabs is active, the demo currently exposes two configured voice buttons:

• JJ-Skånska (default)
• ElevenLabs-Skånska

The default uses ELEVENLABS_JJ_VOICE_ID. The secondary voice uses ELEVENLABS_VOICE_ID.

Files

• server.js: Express server and API routes
• lib/locales.js: locale bundles, route mapping, metadata
• lib/renderPage.js: server-rendered HTML shell per locale
• lib/openaiDialectDemo.js: OpenAI + ElevenLabs pipeline
• public/app.js: tap-to-start/tap-to-stop browser logic
• public/styles.css: styling
• CHANGELOG.md: project history

Changelog

See CHANGELOG.md for the build history from this project thread.