PudimFlix: The open-source player that uses local AI to subtitle and dub live TV on Windows

On October 30, 2026, the "local-first" artificial intelligence ecosystem gained a new practical application focused on media consumption. The developer known as "snoopgameplay" released PudimFlix, an open-source (MIT license) TV player for Windows that provides simultaneous subtitling and dubbing for foreign channels. The fundamental difference compared to established services is that all processing occurs on the user's own hardware, eliminating the need for cloud subscriptions or external API connections.

The software integrates a suite of open models — faster-whisper, Gemma 3 4B, and voice generators like Kokoro and Piper — orchestrated through a PyQt6-based interface and the VLC playback engine. The solution addresses a long-standing accessibility issue: the lack of reliable closed captions in international live broadcasts.

Context and background: The language barrier in live TV

The project's original motivation stemmed from the difficulty of consuming content from foreign news networks and documentaries, such as Bloomberg and ABC News. In live broadcasts (IPTV/HLS), there is rarely a supply of subtitles in multiple languages. Browser-based tools often fail when attempting to capture and process native audio from video players without causing excessive delays or sync issues.

The author began development by creating a local player that downloads HLS video fragments of the channel and hosts a local playlist at 127.0.0.1. This architecture not only enabled the functionality to pause and rewind live TV but also paved the way for direct manipulation of the audio stream.

From there, the developer implemented a three-stage AI pipeline: transcription of the original language, contextual translation, and, optionally, voice synthesis for dubbing with original volume reduction (ducking).

Concrete data and technical architecture

PudimFlix's performance relies on rigorous audio engineering and memory management, developed and tested on an NVIDIA RTX 3050 graphics card with 8 GB of VRAM. The technical pillars of the project include:

• Audio Capture (VLC): The project uses python-vlc. To send audio to the AI without relying on the Windows mixer (speakers), the software uses audio callbacks from the libVLC library (amem). VLC delivers PCM (Pulse-Code Modulation) audio with timestamping (PTS), allowing a 16 kHz mono copy to be sent to the transcription model without losing sync with the video.
• Transcription: It uses faster-whisper with the large-v3-turbo model. The system generates a preview every 1 second (displayed translucently) and consolidates the final subtitle when it detects a pause in speech.
• Translation: After testing with Google Translate and NLLB-200, the final choice was Gemma 3 4B, executed via CTranslate2. The previous sentence is sent as context (marked with the instruction "DO NOT translate") to maintain thematic coherence.
• Dubbing: Handled by local engines Kokoro or Piper, which generate voices in Brazilian Portuguese (pt-BR) inserted at the exact moment of the original video.

During development, engineering hurdles required fine-tuning the libraries:

• VRAM Bottleneck: Running Whisper and Gemma simultaneously on the RTX 3050 made recognition 7 times slower. The solution was to apply int8_float16 quantization to Whisper, keeping peak memory usage around 6.7 GB.
• Gemma Overflow: Initially, running the LLM in int8_float16 caused a bug where the model repeated "our our our our...". Switching to the int8_bfloat16 format resolved the issue.
• Library Crashes: The application would close randomly. Using faulthandler and minidumps, the author discovered a conflict: Qt 6.11, compiled for OpenSSL 3.5, was loading incorrect versions. The problem was solved by switching the Qt TLS layer to the native Windows Schannel.

Impact on the sector and competition

The launch of PudimFlix highlights the rise of "Local-First" AI. Consumers with dedicated graphics cards are increasingly replacing SaaS services with local alternatives, eliminating recurring costs and data privacy concerns.

The impact directly affects streaming companies and startups focused on real-time subtitling and translation. The current market includes significant commercial and hybrid competitors:

• XRAI TV: A hardware solution that acts between the HDMI source and the TV, offering subtitles in over 300 languages for any device (consoles, set-top boxes, streaming).
• Ray Player: A media player focused on AI and an official partner of OpenSubtitles, offering on-demand translation in over 200 languages, combining local and cloud processing.
• DubTab: A service focused on browsers (via extension) and desktop, charging subscriptions based on usage hours for voice translation and dubbing of platforms like Coursera and YouTube.

PudimFlix competes directly with these players by delivering a cost-free solution for those who own the necessary hardware. It also aligns with other open projects, such as Myna Player (currently in Alpha), developed in Rust and Tauri with a focus on the macOS and Windows ecosystem, which also seeks to deliver pre-transcription and contextual translation running locally.

What remains unknown

Despite the functional version documented on the author's GitHub, some technical and market gaps remain:

• Performance on lower-end hardware: The entire project benchmarking was done on an 8 GB NVIDIA GPU. There is no consolidated data on PudimFlix's behavior on integrated video processors, AMD cards, or GPUs with 4 GB of VRAM or less.
• Residual latency: Transcription depends on natural pauses in speech to finalize the subtitle. It is unclear how significant the perceived delay is in programs with continuous or overlapping speech (such as political debates or frantic sports broadcasts).
• Portability to other systems: Although it uses cross-platform technologies like Python, PyQt6, and VLC, the binaries and packaging solutions (PyInstaller, Inno Setup) and fixes (Schannel) are strictly focused on Windows. There is no confirmation of future support for Linux or macOS.

What to expect next

With the source code released under an MIT license, PudimFlix is now entering a phase of community validation. Likely next steps include:

• Hardware stress tests: Independent developers are expected to compile and run the player on AMD and Intel ecosystems, potentially adapting the code to use ROCm or OpenVINO to democratize access for those without NVIDIA cards.
• Latency optimization: The open-source AI community will likely propose pull requests focused on predictive translation algorithms to reduce the time between speech and subtitle rendering.
• Third-party list integrations: As the player supports M3U files and already aggregates open catalogs (iptv-org, Pluto, Samsung TV Plus), the emergence of community repositories optimized specifically for the player's dubbing tests is expected.

Frequently Asked Questions

1. Is PudimFlix paid or does it require a monthly subscription? No. The software is open-source (MIT) and completely free. The AI models (Whisper, Gemma 3, Kokoro) run locally on the user's machine, eliminating the need for paid API keys or cloud subscriptions.

2. What hardware is needed to run PudimFlix? The current version was optimized and tested on an NVIDIA RTX 3050 graphics card with 8 GB of VRAM, using memory quantization to avoid overloading. Lower-end GPU models or those without CUDA acceleration may face extreme performance bottlenecks.

3. Does the player work with any video or only live channels? The system was designed for HLS transmissions (live IPTV channels) with a playlist served locally. It supports adding custom M3U lists, but its architectural focus is handling real-time streaming where subtitles do not pre-exist.