mrtdlgc

@mihrac evet abi geldi. subtitle foundry'i de baştan sonra test etmiş oldum. normalde en az 10 saat sürecek iş filmi de izlemeyle birlikte 2 saatte bitti mrtdlgc/subtitle-foundry

My Scribe v2 subtitle workflow: full-recording FLAC, native word timing and manual review I've been using ElevenLabs Scribe v2 as the transcription stage of a subtitle workflow, with Turkish-to-English and Ukrainian-to-English film testing. I'm a former professional translator/subtitler/interpreter and the developer of the open-source implementation, Subtitle Foundry. Here is how the Scribe path is set up: 1. Prepare one continuous 16 kHz mono FLAC. In the default full-recording mode, I retain music and silence rather than cutting the recording into speech-only segments. Optional short chunks are a separate processing choice. 2. Use Scribe's native word timestamps. This route doesn't run a second local speech-detection/alignment model. I keep the original response so the words and timing evidence remain inspectable. 3. Build cues from actual words and speaker turns. Speaker changes help split the subtitles. Anonymous speaker labels remain labels; the pipeline doesn't claim to identify the people speaking. 4. Flag timing problems for review. Timestamp anomalies and unusually long cues need inspection. I don't want postprocessing to hide a timing issue just because the exported file looks tidy. 5. Keep transcription separate from translation. The downstream translation uses dialogue context, while retaining source timing evidence. There is also an exact-timing mode for existing cues. A fluent translation still needs checking against what was said. 6. Finish with playback and waveform review. Wording and timing corrections are saved, and the result exports to SRT or ASS. Manual edits don't trigger another Scribe request; matching saved responses can be reused when resuming work. I haven't published a controlled provider comparison. These are the settings and implementation choices behind my current workflow, rather than a claim that the same configuration wins on every recording. Implementation and setup documentation: mrtdlgc/subtitle-foundry I'm not making money from this AGPL-3.0 project, and I'll keep maintaining it. The purpose is to bring state-of-the-art (SOTA) tools into a reviewable subtitle workflow for material that would otherwise fail to reach a global audience. Provider usage costs remain separate from the free app. I used Codex while building it and preparing this write-up.

I made Subtitle Foundry to help films without usable subtitles reach a global audience I'm a former professional translator, subtitler and interpreter. After about 15 years away from development, I came back to building and made Subtitle Foundry. It's a free, open-source subtitle workbench. I'm not making money from it, and I'll keep maintaining it. The purpose is to make subtitles for films, documentaries and other material that would otherwise fail to reach a global audience. You can start with a local video, a YouTube URL or an existing subtitle file. Transcription, alignment, context-aware translation and review are separate stages, with choices of providers and supported processing options. It uses state-of-the-art (SOTA) speech and language tools, and includes a video/waveform subtitle editor so a person can listen, correct wording and timing, and export SRT or ASS. I've tested Turkish-to-English and Ukrainian-to-English, mainly with ElevenLabs Scribe v2 and Codex. You can run the app locally or deploy it to your own server. The app is AGPL-3.0; cloud model providers may charge for usage or consume your existing subscription allowance. I used Codex while building it. Working on this brought me back to JavaScript/TypeScript and Python, and taught me a lot about audio processing and model hallucinations. Source and setup guides: mrtdlgc/subtitle-foundry

Subtitle Foundry: AGPL-3.0 subtitle transcription, translation and waveform review I'm sharing Subtitle Foundry, a subtitle workbench I've released under AGPL-3.0-only. Source and installation guides: mrtdlgc/subtitle-foundry I'm the developer and a former professional translator, subtitler and interpreter. I'm not making money from the project, and I'll keep maintaining it. The purpose is to help films, documentaries and other material reach a global audience when usable subtitles would otherwise be missing. The application separates media preparation, transcription, alignment, translation and review. You can select supported providers/options, reuse existing subtitles, inspect intermediate artifacts and resume matching completed work. The editor includes video playback, a waveform, editable subtitles, persistent corrections and SRT/ASS export. It runs locally or on your own server, with Docker and native setup documented in the repository. I'm building around state-of-the-art (SOTA) speech and language tools while keeping the stages replaceable. My film testing has mainly used ElevenLabs Scribe v2 and Codex for Turkish-to-English and Ukrainian-to-English. Human listening and correction remain part of the workflow; I haven't published a comparative benchmark for the full pipeline. The open-source application and its model providers are separate. The default ASR and translation workflows call cloud services, including proprietary providers, and can incur provider charges or consume an existing subscription allowance. The app has no paid feature tier, but the default workflow is not a fully offline or entirely FOSS model stack. Optional local ASR is available; its additional dependencies are documented rather than bundled in the default image. Personal installations support subscription connections for Codex, Claude Code and Grok Build where available, alongside explicit API provider choices. You can also start from an existing SRT and use the translation/review workflow without a new ASR run. I used Codex while building the project. Contributions around usability, subtitle timing, language coverage and reproducible evaluation would be welcome. If you try it, starting with a short clip is a useful way to inspect the workflow and provider usage before processing a full film.

Subtitle Foundry - a modular subtitle workbench to help films reach audiences across languages I'm a former professional translator, subtitler and interpreter. After about 15 years away from development, I started building a subtitle pipeline around the workflow I wanted to use. I called it Subtitle Foundry. I'm not making money from it, and I'll keep maintaining it. The purpose is to make subtitles for films, documentaries and other material that would otherwise fail to reach a global audience. Repo: mrtdlgc/subtitle-foundry The app combines state-of-the-art (SOTA) speech recognition and language tools with separate stages for transcription, alignment, context-aware translation and review. You can choose providers and supported processing options, reuse existing subtitles, and inspect the intermediate work. You can run it locally or deploy it to a VPS, start with a local video or YouTube URL, or translate an existing SRT. My current workflow uses ElevenLabs Scribe v2 and Codex; I've tested Turkish-to-English and Ukrainian-to-English. Other transcription and translation connections are available, including personal subscription connections for Claude Code and Grok Build. After generation, there's a subtitle editor with video playback and an audio waveform, so you can check wording and timing, save corrections, and export SRT or ASS. Context helps with difficult dialogue, but I still want a person to be able to listen and correct the result. Building this taught me a lot about model hallucinations and audio processing, and brought me back to JavaScript/TypeScript and Python. One practical priority became keeping completed provider responses and checkpoints, so a review or an interrupted job doesn't mean starting the whole process again. The source is AGPL-3.0 and has setup guides. The app is free; chosen model providers can charge you or use your subscription allowance. The SOTA description refers to the tools used in the workflow; I haven't published a benchmark ranking the whole pipeline. I'd appreciate feedback on the first-run experience and the subtitle editor, especially from people working with language pairs that get little attention.

My subtitle workflow: full-film ASR, context-aware translation and waveform review in Subtitle Foundry Disclosure: I built Subtitle Foundry, a free, open-source subtitle workbench. I'm not making money from it, and I'll keep maintaining it. My purpose is to help films, documentaries and other material reach a global audience when usable subtitles would otherwise be missing. I'm a former professional translator, subtitler and interpreter. The project brings state-of-the-art (SOTA) speech recognition and language tools into a configurable workflow, with transcription, alignment, translation and review kept as separate stages. Repo and setup guide: mrtdlgc/subtitle-foundry Here is the workflow I've been using for Turkish-to-English and Ukrainian-to-English: 1. Start with the best available source. Choose a local video or YouTube URL. The app checks for suitable existing subtitles before generating a new transcription. An existing SRT can also go straight to the Translations desk. 2. Transcribe when needed. My current default is ElevenLabs Scribe v2 with the complete recording and native word timestamps. Other ASR providers and speech-aware chunks are available, so you can choose what suits the source and provider limits. 3. Translate with context. I use Codex with my existing subscription on a personal installation. The translation can preserve cue timing, or the editorial mode can form target-language cues using source timing evidence. Claude Code, Grok Build and explicitly selected API providers are also supported connections; my film testing has mainly used Codex. 4. Watch and correct. Review against the video and audio waveform, fix wording or timing, save corrections, and export SRT or ASS. The intermediate results remain available, and manual edits don't make another model request. For a first try, install with the README's Docker/local instructions, add your provider credentials in Connections, and use a short clip or an existing SRT. The default cloud workflow doesn't need a local GPU. You can also deploy the app to a VPS using the deployment guide. A few practical limits: dialogue context can help with ambiguity, but a fluent translation still needs checking against the audio. I haven't published a benchmark comparing the complete pipeline against other tools. The software is AGPL-3.0; any model API charges or subscription usage go to the providers, not to me. Running the app locally doesn't make cloud ASR and translation offline. The part I most want to keep improving is the path from generated text to subtitles someone has actually listened to, corrected and can share with another audience.