Clarity for AI buildersSponsored by AI Routing
adoption

Nokia N95 8GB runs a voice-enabled ChatGPT client through a Linux bridge

Nokia N95 8GB runs a voice-enabled ChatGPT client through a Linux bridge

Conor O’Neill reports running a text-and-voice ChatGPT client on a Nokia N95 8GB, with a Linux computer handling API access.

In a personal project write-up, O’Neill says Codex produced a working first text version and that voice support followed in less than an hour. He reports a 19KB JAR for the initial text application. The result is a handset demonstration, not a new official ChatGPT release or evidence that OpenAI models run locally on the phone.

The post includes a technical account explicitly identified as written by ChatGPT. According to that account, the Java ME application exchanges messages with a small Python service over Wi-Fi. The service handles modern HTTPS connections, OpenAI API credentials and audio conversion. O’Neill says the phone lacks the certificates and security capabilities needed for direct access; the technical account adds that direct connectivity from a patched handset was not established.

That division of work is central to the experiment. The N95 supplies the keypad, screen, microphone and speaker, while the bridge adapts remote services to the older device. It demonstrates that the handset can serve as an interface to current AI services without demonstrating that its operating system can independently support those services’ security requirements.

Text messages and voice turns share conversation context. The bridge uses OpenAI’s Responses API, returns a response ID to the phone and supplies it as previous_response_id on the next request. The application targets MIDP 2.0 and CLDC 1.0, uses standard Java ME text-entry controls and stores settings in the Record Management System. Conversation text and the current response ID remain in memory.

For voice input, users hold the centre navigation button or the 5 key, speak and release the button. A Record / Stop command offers an alternative. Recordings are limited to 30 seconds and two MiB. The bridge converts incoming audio to mono 16 kHz WAV with FFmpeg, requests transcription, sends the recognized text to the conversation model and requests an MP3 version of the answer.

Both recognized words and the answer appear on screen.

This is turn-based voice, not a streaming Realtime implementation. Each processing stage completes before the next begins, and the phone downloads the complete MP3 before playing it. The text answer appears before speech preparation finishes and remains available if speech generation or playback fails. The distinction matters: the demonstrated capability is a sequence of recorded requests and spoken replies, rather than continuously streaming conversation.

The write-up also describes hardware-specific debugging. A microphone permission dialog initially hid the voice screen, causing the application to cancel recording. The fix separated microphone initialization from active recording and allowed the Stop command to finish a recording when the permission dialog consumed the original key release. Permission-setting options may vary with firmware and application trust.

The reported testing combines actual handset use with narrower checks. O’Neill tested text, recording, transcription and spoken replies on the N95 8GB. Python integration tests exercised local HTTP requests and FFmpeg conversion with simulated OpenAI responses; a separate live speech-generation request produced a valid MP3. The included macOS route was not tested, and the working handset recording’s exact encoding was not separately verified.

There are important deployment limits. Phone-to-bridge traffic uses plain HTTP, so the documented setup is intended for a trusted local network. The bridge sends relevant content to OpenAI over HTTPS, and API stages incur separate usage charges. Clearing the local conversation does not delete responses stored for conversation continuation. A proposed ESP32 companion remains unimplemented, so the demonstrated setup still depends on the computer-hosted bridge.