AiSrt: translate, extract and transcribe your subtitles with AI

aisrt "film.mkv"

That's it. This line gives you a subtitle in your default language, whatever the video.

The tool

What AiSrt is for

Four jobs in a single little program, and AiSrt chooses on its own which one your video needs.

A translation in progress: each batch is sent to the chosen service, the bar shows the remaining time.

Why you’ll love it

Installation

Ready in three steps

# Windows (PowerShell)
.\install.ps1

# Linux et macOS
sudo ./install.sh

aisrt setup

No key yet? At OpenAI, it takes two minutes: create an account, add a few dollars of credit, create a key on the API keys page and paste it when aisrt setup asks for it. A full movie costs a few cents. And if you don't want to pay anything at all, Ollama runs the AI on your machine, with no account or key.

First steps

# Une vidéo : sous-titres dans votre langue par défaut
aisrt "film.mkv"

# Plusieurs langues, et une copie de la vidéo qui les embarque
aisrt "film.mkv" -t fr,es,pt-BR --mux

# Un fichier de sous-titres : la langue source est détectée toute seule
aisrt "film.srt" -t de

# Une série entière, sous-dossiers compris
aisrt "D:\Series\The Wire" -r -t fr

# Surveiller un dossier et traduire tout ce qui y arrive
aisrt "/media/incoming" --watch -t fr

What you need

Extraction and transcription require neither an account nor a key, and send nothing anywhere.

Going further

The technical side

Everything that follows is optional: the default settings are suitable for most uses. This is where I explain how AiSrt works and how to adapt it to your needs.

How AiSrt chooses its path

When faced with a video, AiSrt always takes the least expensive path:

1. a track already exists in the requested language: it extracts it, nothing more ;
2. a text track exists in another language: it extracts and translates it (the language from -s first, then English, then the others) ;
3. there is only an image track: it reads it with OCR, then translates it ;
4. there are no subtitles: it transcribes the audio, then translates it.

--extract and --transcribe force a specific path. The source language is inferred from the name (film.en.srt), the track label, Whisper, or the model taking a glance at the text: -s only serves to force it.

Translation services

AiSrt talks directly to Anthropic (Claude), Google (Gemini), xAI (Grok), OpenAI and Ollama, as well as everyone who speaks the same language as OpenAI: Mistral, DeepSeek, Groq, OpenRouter, Together, Fireworks, Cerebras, Azure OpenAI, LM Studio, llama.cpp and vLLM.

For services whose catalog changes every month, I deliberately don't lock in any model: AiSrt reads the service's list, lets you choose and keeps your choice. aisrt --providers tells you which ones are ready.

aisrt "film.srt" -t fr -p mistral     # pour cette commande seulement
aisrt config set provider mistral     # pour toutes les suivantes

Before a paid translation, AiSrt estimates the cost and asks for confirmation above 0.50 dollar (adjustable). At the end, it displays the tokens (the units of text charged by the service) actually consumed.

Locally with Ollama

ollama pull gpt-oss:20b
aisrt "film.srt" -t fr -p ollama

I tested several models, three times each. gpt-oss:20b, the one I ship by default, never lost a single block, but it weighs about 14 GB and takes its time. llama3.1:8b goes about ten times faster for an almost equally clean result. qwen2.5, for its part, slipped comments in Chinese into the middle of the subtitles, three times out of three: avoid it. Locally, it's free but slower, and the quality remains a notch below the best online services.

Glossary and reading speed

# aisrt-glossary.txt, dans le dossier de la série
Walter White
Heisenberg
the Cousins = les Cousins

A term on its own is kept as is, a term followed by = is always translated the same way. The file is found automatically in the episode, season or series folder, and only the terms present in each batch are sent.

A translation is often longer than the original. After each file, AiSrt counts the subtitles displayed for too little time for their length (above 20 characters per second; Netflix aims for 17). With --condense, the model shortens them, and any rewrite that would not be shorter is thrown out. All the texts sent to the model can be modified in the configuration.

Formats and files produced

• a .srt gives a .srt, an .ass gives an .ass with its styles (--to-srt for a simple .srt) ;
• .vtt, .sub, .sbv, .smi and .ttml are converted with ffmpeg ;
• the files are named according to the language (film.fr.srt, film.pt-BR.srt), as Plex and Jellyfin expect ;
• --mux writes film.aisrt.mkv, a copy of the video with the new tracks labelled ;
• an existing file is always backed up as .bak.srt before being replaced, and an interrupted translation resumes where it left off.

Blu-ray subtitles

The PGS tracks on Blu-rays store a picture of the words, not the words. Two engines can read them:

Tesseract (by default, local and free), an optical character recognition (OCR) program, it reads the text in an image. It's the only dependency AiSrt doesn't download, because it installs differently on every system:

winget install tesseract-ocr.tesseract   # Windows
sudo apt install tesseract-ocr          # Linux
brew install tesseract                  # macOS

The recognized text is then reread by your AI service, with a strict instruction: correct reading errors (a slash instead of an I, an "rn" read as an "m"), without translating or rephrasing anything. Out of 100 test subtitles, 14 were corrected and none were moved. I prefer a visible typo to a perfect line stuck in the wrong place, so any doubtful answer keeps the raw text.

A vision model reads the images directly, much better with stylized fonts, without installing anything:

aisrt config set ocr-engine vision

A feature-length film contains about a thousand images: AiSrt announces the number of requests before starting, and rereads frame by frame any response that does not return exactly one text per image. Fades, where the disc displays the same image with a different transparency, no longer cut a line into pieces: AiSrt keeps the most readable rendering.

The transcription

When a video has no subtitles, AiSrt uses whisper.cpp, the optimized version of OpenAI's Whisper speech recognition model. The spoken language is detected, and under Windows an NVIDIA card is used when one is present.

aisrt --transcribe "conference.mp4" -tt fr   # transcrire, puis traduire en français

When faced with French audio, the .en models do not crash: they invent very confidently worded English text. AiSrt therefore rejects this combination. Only speech is transcribed, music and noises produce no subtitles.

Checks before sending

Before the first paid call, AiSrt checks that the service responds: invalid key, account without credit, unknown model, local server turned off. Each case has its message and the setting to correct. It also checks the nature of each file: nothing binary is sent to the translation service.

Automate

aisrt "film.mkv" -t fr --non-interactive
aisrt --extract "episode.mkv" --stream 2 --translate-to fr,es --non-interactive

With --non-interactive, each question uses its default value and the log says which one. An explicitly given value is never ignored: asking for track 7 of a file that only has two stops everything and lists the available tracks. Exit codes: 0 success (including nothing to do), 1 at least one file failed, 2 cancellation by Ctrl+C.

Configuration

aisrt config list                     # les réglages en vigueur, clés masquées
aisrt config set target pt-BR         # langue par défaut
aisrt config set Translation:ConfirmCostAbove 2
aisrt config path                     # où chaque réglage est lu

The settings live in your user folder (%APPDATA%\AiSrt on Windows, ~/.config/aisrt elsewhere). A key can also come from an environment variable under the service's usual name (OPENAI_API_KEY, MISTRAL_API_KEY...), and every configuration key can be overridden with a variable prefixed AISRT_, handy in a container.

What does not work yet

Image subtitles from DVDs and DTT are not decoded: they are read using information stored elsewhere in the file, that is another project. AiSrt detects them and warns you instead of silently ignoring them.