GPT Subtitler

1. Sign in to your account or create a new account if you don't have one.
2. Click on "Translate a new subtitle" to create a new translation project.
3. Upload your subtitle file in the SRT format or paste the subtitle content directly into the editor.
4. Click on "Create" to start the translation process.
5. If you have an OpenAI/Claude/Gemini API key, you can enter it in the settings. Otherwise, you need to use your account token for translations.
6. Configure the translation settings, such as source and target languages, start index, prompt, batch size and model.
7. Click on "Translate" to begin the translation.
8. The translation process will start, and you can monitor the progress in real-time. The translated subtitles will be displayed in the "Translation Messages" section.
9. You can stop the translation process at any time and resume it later using the "start index" option.
10. Once the translation is complete, you can download the translated subtitles in your desired format.

GPT Subtitler supports the SRT (SubRip Text) file format for subtitle uploads. If your subtitles are in a different format, you can convert your subtitles file at subtitletools .

GPT Subtitler offers different pricing plans to suit your needs. We have a free subscription option, where you can receive a certain amount of Tokens daily, which can be used to translate a ~30 minutes video using the GPT-4o-mini model. For higher usage and advanced features, we offer Basic and Premium subscription plans. Please visit our pricing page for more details on the available plans and their pricing.

You can use OpenAI's Whisper to transcribe the video, and then use the transcribed subtitle file as the input for the translation. If you just want to transcribe the video file, you can download the tool from this GitHub repository vibe. If you want to download a Youtube video, transcribe it using Whisper and translate it using GPT in one go, I recommend using this tool I've open-sourced on my GitHub repository GPT_subtitles. This repo is the inspiration for this project, and if you find it useful, please consider leaving a star on the repo ⭐. Thank you!

The token cost for translations is calculated based on the number of input and output tokens required to process the subtitle content. OpenAI charges different prices for input and output tokens, and the prices vary depending on the model used.

In GPT Subtitler, we use the token consumption of the $0.5/1M Tokens as baseline. Then, we calculate the token consumption based on the actual token usage and the model used, proportionally.

Please note that while the LMSYS Score represents the general capability of the model, the actual performance for translation tasks may vary for some language pairs. But generally, Gemini models are the best choice for non-English translation.

• The Context Window represents the maximum number of tokens the model can process in a single request.
• Max Generation is the maximum number of tokens the model can generate in a single response.
• Output Speed is a relative measure of how quickly the model generates responses.
• The LMSYS Score is a benchmark score from lmsys.org, indicating the model's general performance across various tasks. (As of 2025-02-06)

The total token cost for a translation request is calculated by multiplying the number of input tokens by the input token price and multiplier, and the number of output tokens by the output token price and multiplier. The sum of these costs gives the final token price for the translation.

You can view the estimated token usage and price for each translation request in the "Translation Cost" section when creating a new translation project. If you have a paid subscription plan, you may have a daily token allowance that you can use for translations without incurring additional costs.

Note: Basic plan users get a 5% discount on token usage, and Premium plan users get a 10% discount on token usage.

Example 1: Using Claude-Haiku model with 1000 input tokens and 2000 output tokens:

Input token = 1000 * 0.5 = 500
Output token = 2000 * 2.5 = 5000
Total token = 500 + 5000 = 5500

Example 2: Using GPT-3.5-TURBO model with 1000 input tokens and 2000 output tokens:

Input token = 1000 * 1.0 = 1000
Output token = 2000 * 3.0 = 6000
Total token = 1000 + 6000 = 7000

Example 3: Premium plan users using Claude-Haiku model with 1000 input tokens and 2000 output tokens:

Input token = 1000 * 0.5 = 500
Output token = 2000 * 2.5 = 5000
Total token = (500 + 5000)*0.9 = 4950

Cache Pricing Information

We use Anthropic's cache service for Claude models and cache the system prompt for each translation request, this can significantly reduce the token usage and cost for subsequent translations if you use the same model and settings and have a long additional context and system prompt. The following table shows the pricing information for cache operations:

For more information about prompt caching, see Anthropic's blog post on prompt caching

Our platform offers several transcription models with different cost structures and performance characteristics. The table below compares these models based on token cost per minute, processing speed, word error rate, and language support.

Important Notes on Transcription Models

• Token cost is calculated based on the duration of the audio file, with different models having different token rates per minute.
• Word Error Rate (WER) is a measure of transcription accuracy - lower percentages indicate better performance.
• Processing speed indicates how quickly the model can transcribe audio relative to its duration.
• Number of languages supported by each model. Some models are optimized for specific languages.
• AssemblyAI models are billed per second with rounding to the nearest second.
• FAL AI models (Whisper, Wizper) are billed by exact duration with no rounding.

For general use, we recommend Whisper Large v3 for the best balance of accuracy and cost. For English-only content requiring highest accuracy, AssemblyAI Slam-1 is recommended. For cost-sensitive applications, AssemblyAI Nano offers good performance at a lower price point.

Yes, you can choose to use your own OpenAI/Claude/Gemini API key for translation. If you provide an API key, the server will use it for translation, and your account Token balance and usage will not be affected. If you use a paid subscription plan, you can also use the translation service without providing an API key and without using Token balance, until your Token quota is exhausted.

To get an OpenAI API key, you need to register an account on the OpenAI website at openai signup and openai api-keys . Once you have the API key, you can use it in the translation settings to enable the use of the OpenAI API for translations. This website will not store your API key or use it in any other way than for the translation process. It is recommended to revoke your API key once in a while to ensure the security of your account.

To get a free Gemini API key, visit https://aistudio.google.com/app/apikey. Follow the instructions to generate your API key. Once you have the API key, you can use it in the translation settings to enable the use of the Gemini API for translations.

1. Sign in to your account.
2. Navigate to the "Settings" page.
3. On the settings page, you will find information about your token balance, subscription tier, and usage details.
4. You can also view your token usage history and transaction logs on the same page.

The Whisper transcription feature is a new addition to our service that allows you to automatically transcribe audio and video files. It uses OpenAI's Whisper model to generate accurate transcriptions in multiple languages. This feature is particularly useful for creating subtitles for videos, transcribing interviews, or converting speech to text for various purposes. You can access this feature in the 'Audio Transcriber' section of our website.

1. Go to the 'Audio Transcriber' section of the website.Try it now!
2. Upload your audio or video file, or provide a YouTube URL.
3. Select the language of the audio content (if known).
4. Optionally, add any specific instructions or context in the 'Prompt' field.
5. Click the 'Start Transcription' button to start the process.
6. Once completed, you can view, edit, and download the transcription, or create a new subtitle project with the result.

The batch translation feature allows you to translate multiple subtitle files in one project. Here are the key points:

• Maximum 10 files per batch project
• Free users: translate up to 2 files simultaneously
• Basic/Premium users: translate up to 5 files simultaneously

• Create a new batch project
• Upload your SRT files (drag & drop supported)
• Select files to translate
• Configure translation settings once for all files
• Monitor progress and download completed translations

• Avoid using free model to translate too many files at once as it may cause rate limit issue
• Preview files before translation
• Download individual files or all completed translations at once
• Stop and resume translations as needed

GPT Subtitler supports translation between all languages supported by the language models, including English, Chinese, Spanish, Japanese, Korean, and more. It can even translate some fictional languages, such as Elvish or Valyrian. When creating a new translation project, you can select the source and target languages from the provided dropdown menus. If your desired language is not listed, you can select "Other" and manually enter the language. Please note that the quality of translation may vary depending on the language pair and the model used. Currently, this project is optimized for English to Chinese translation, other language pairs may have different effects.

The backbone of the translation service of our website is available in GitHub GPT_subtitles , but in short, the translations are processed by sending the subtitle content in batches to the OpenAI/Claude/Gemini API for processing. The prompt is carefully crafted to guide the model to generate accurate translations. We also utilize additional context provided by the user and few-shot examples to improve the translation quality. The API generates the translated text based on the input content and the specified settings, such as source and target languages, prompt, and model. The translated text is then displayed in real-time on the website, allowing you to monitor the progress and download the translated subtitles once the process is complete.

Few-shot examples are a small set of example translations provided to the AI model to guide and improve the quality of the generated translations. By including a few relevant examples of how specific phrases or sentences should be translated from the source language to the target language, the model can better understand the desired translation style, tone, and context. This approach helps to produce more accurate and contextually appropriate translations, especially for domain-specific or idiomatic expressions. The minimal required format for few-shot examples is as follows:

[
  {
    "input": {
      "current_batch_subtitles": [
        {
          "index": 1,
          "original_text": "This is a sample subtitle."
        }
      ]
    },
    "output": {
      "current_batch_subtitles_translation": [
        {
          "index": 1,
          "original_text": "This is a sample subtitle.",
          "first_translation": "这是一个示例字幕。",
          "translation": "这是一个示例字幕。"
        }
      ]
    }
  }
]

Optionally, you can include 'previous_batch_subtitles' and 'next_batch_subtitles' fields to provide additional context for the translations. These fields should be placed inside the 'input' object, alongside the 'current_batch_subtitles' field. The structure for these optional fields is as follows:

{
  "previous_batch_subtitles": [
    {
      "index": 1,
      "original_text": "Previous subtitle 1",
      "first_translation": "之前的字幕1",
      "translation": "之前的字幕1"
    }
  ],
  "next_batch_subtitles": [
    {
      "index": 3,
      "original_text": "Next subtitle 1"
    }
  ]
}

Prompt for generating few-shot examples. After generating the examples, copy the final JSON array into the few-shot examples section. This should improve the accuracy of the translation and reduce the error rate of the model:

Source Language: (fill in here)
Target Language:  (fill in here)
I need few-shot examples for translating subtitles from [Source Language] to [Target Language]. Please provide 5-7 examples that showcase common subtitle structures, idiomatic expressions, and cultural references. Each example should include the original text in [Source Language], its translation in [Target Language], and any relevant context or explanation. The examples should be formatted as follows:

{
  "input": {
    "current_batch_subtitles": [
      {
        "index": 1,
        "original_text": "[Source Language Text]"
      }
    ]
  },
  "output": {
    "current_batch_subtitles_translation": [
      {
        "index": 1,
        "original_text": "[Source Language Text]",
        "translation": "[Target Language Translation]"
      }
    ]
  }
}

Please ensure the examples cover a range of scenarios typical in subtitle translation.

After providing the individual examples, please combine all the examples into a single JSON array structure, like this:

[
  {
    "input": { ... },
    "output": { ... }
  },
  {
    "input": { ... },
    "output": { ... }
  },
  ...
]

This combined JSON structure should include all the examples you've created, making it easy to use as a complete set of few-shot examples for the translation model.

Based on experience, Gemini-pro generally offers the best performance for translation tasks. However, the optimal model can vary depending on the specific language pair. If Gemini-pro's performance doesn't meet your expectations, try the following models in this order:

1. Gemini-pro
2. Claude-3.5-sonnet
3. GPT-4o
4. GPT-4o-mini
5. Gemini-flash
6. Claude-haiku

It's recommended to test these models with a smaller batch of text to determine which works best for your specific use case. Note that DeepSeek models are exceptionally good and cost-effective for Chinese translation but may not perform as well for other languages.

Recommended settings:

1. Batch Size: Use a higher batch size to provide more context for the translation.
2. Enable Context Memory: This helps the model better understand the video context and produce more accurate translations, although it may slightly increase token usage.
3. Fluent Merging: Beneficial when using the Whisper model to generate subtitles. It helps handle cases where sentences are broken across multiple lines, creating more natural translations.
4. Few-shot examples: These can significantly improve translation accuracy by providing guidance to the model. As these are language-specific, you may need to create custom examples for your language pair.

A setting incorporating many of these features is available at: https://gptsubtitler.com/settings/8dd9a960-cb30-4141-8b23-d3944d97773e. You can access and modify it to fit your specific needs.

In the settings modal, click on the 'Share' button next to your setting. You can then review and edit your setting, after publish the setting, you can use "copy link" to get a shareable link.

Yes! You can access the translation settings by opening link shared by others, bookmark ones you like, and load them into your translation settings.

Go to your account settings. You'll find an option to edit your username next to your username. If you haven't set one, a random username will be assigned which you can then change.

Custom instructions in the Additional Context field give the AI model specific guidance on how to translate your subtitles. Here are some tips for writing effective instructions:

1. Be specific and clear: Instead of "make it sound good", try "translate using casual language appropriate for a teenage audience".

2. Focus on one aspect at a time: Separate formatting instructions ("preserve line breaks") from style instructions ("use formal language").

3. Provide examples when possible: If you want specific terms translated in a certain way, include examples.

4. Consider cultural context: If the content has cultural references, specify how to handle them ("explain Japanese cultural references briefly in parentheses").

5. Use the Instruction Library: Access pre-made instruction templates by clicking the "Instruction Library" button below the Additional Context field.

Remember that instructions should be in the target language or English, as they're meant to guide the AI, not to be translated themselves.

If you have any further questions, feedback, or need assistance, please feel free to contact me at support@gptsubtitler.com . I will be happy to help you with any queries or issues you may have.

I'm sorry that we didn't meet your expectations. We support refunds within 7 days after payment. Please contact me through the email you used when registering (support@gptsubtitler.com), and I will do my best to resolve your issue.