AI-powered subtitle generator that combines ElevenLabs Scribe for precise character-level timestamps with OpenAI LLMs for intelligent sentence segmentation — producing accurately timed SRT subtitles from audio files.
Traditional speech-to-text tools either give you a wall of text with no timing, or break lines at awkward points. ExactSub solves this by splitting the work:
- ElevenLabs Scribe v2 handles transcription with character-level timestamps
- OpenAI GPT handles semantic segmentation — breaking text into natural, readable subtitle lines
- A custom alignment algorithm merges the two, mapping each segmented line back to precise timestamps
The result: subtitles that are both accurately timed and naturally segmented.
- Accurate timestamps — Character-level alignment from ElevenLabs Scribe v2
- Semantic segmentation — LLM-powered line breaking that respects meaning and reading flow
- Multiple subtitle styles — YouTube (full sentences, 10–30 chars/line) or TikTok (short bursts, 3–8 chars/line)
- Multi-language support — Chinese, English, Japanese, Korean, Cantonese, or auto-detect
- Speaker diarization — Identify and annotate different speakers in the audio
- Keyterms — Bias transcription towards specific terms for better recognition accuracy
- Text correction — Optional LLM pass to fix transcription errors
- Translation — Translate subtitles into Traditional Chinese, English, Japanese, or Malay
- Simplified → Traditional Chinese — Lightweight OpenCC conversion (no LLM needed)
- Keyword rules — Define glossary replacements (e.g.,
OpenAI=>OpenAI) or terms to preserve - Punctuation cleaning — Remove leading dashes, trailing commas/periods from subtitle lines
- Cost tracking — Real-time cost estimation for both ElevenLabs and OpenAI API usage
- Transcription model selection — Choose between Scribe v2 (recommended) and Scribe v1
- Multiple LLM models — GPT-4.1, GPT-4.1 mini, GPT-5.2, GPT-5-mini with configurable reasoning effort
Audio File (mp3/wav/m4a/flac/ogg)
│
▼
┌─────────────────────┐
│ ElevenLabs Scribe │ → Transcription with character-level timestamps
└─────────┬───────────┘
│
▼
┌─────────────────────┐
│ OpenAI LLM │ → Semantic segmentation into subtitle lines
└─────────┬───────────┘
│
▼
┌─────────────────────┐
│ Alignment Engine │ → Maps each line to precise start/end times
└─────────┬───────────┘
│
▼
┌─────────────────────┐
│ Post-Processing │ → Correction, translation, OpenCC, cleanup
└─────────┬───────────┘
│
▼
SRT File
- Python 3.9+
- An ElevenLabs API key
- An OpenAI API key
git clone https://github.com/leonchanwy/ExactSub.git
cd ExactSub
pip install -r requirements.txtstreamlit run app.pyThen open the URL shown in the terminal (default: http://localhost:8501).
- Enter API keys in the sidebar (ElevenLabs + OpenAI)
- Configure settings:
- Choose subtitle style (YouTube / TikTok)
- Set max characters per line
- Select OpenAI model
- Choose audio language (or auto-detect)
- Upload an audio file (mp3, wav, m4a, flac, or ogg)
- Click "開始生成字幕" (Start generating subtitles)
- Download the generated
.srtfile
| Option | Description |
|---|---|
| Transcription model | Choose Scribe v2 (recommended) or Scribe v1 |
| Speaker diarization | Annotate different speakers in multi-person audio |
| Custom segmentation prompt | Additional instructions for the LLM segmenter |
| Punctuation cleaning | Strip leading dashes and trailing punctuation |
| Text correction | LLM pass to fix transcription typos |
| Translation | Translate to another language via LLM |
| OpenCC conversion | Lightweight Simplified → Traditional Chinese (no LLM) |
| Keyword rules | source=>target replacements or keyterms to preserve (also sent to ElevenLabs for better recognition) |
| Reasoning effort (GPT-5) | Control thinking depth: none → xhigh |
| Retry settings | Configure upload timeout and retry behavior |
# Replacement rules (one per line)
王小明=>小明
OpenAI=>OpenAI
# Preserve terms (no arrow, kept as-is)
Hokage
The app tracks API costs in real time:
| Service | Pricing |
|---|---|
| ElevenLabs Scribe | $0.008 / minute of audio |
| OpenAI GPT-4.1 | $2.00 / 1M input tokens, $8.00 / 1M output tokens |
| OpenAI GPT-5.2 | $1.75 / 1M input tokens, $14.00 / 1M output tokens |
| OpenAI GPT-5-mini | $0.25 / 1M input tokens, $2.00 / 1M output tokens |
MIT
結合 ElevenLabs Scribe 的字元級精準時間戳與 OpenAI LLM 的語意斷句能力，從音訊檔案自動生成高品質 SRT 字幕。
傳統語音轉文字工具要嘛只給你一整段沒有時間軸的文字，要嘛在不自然的地方斷行。ExactSub 把工作分成兩步：
- ElevenLabs Scribe v2 負責轉錄，提供字元級時間戳
- OpenAI GPT 負責語意斷句，根據語意和語氣自然換行
- 自製的對齊演算法將兩者合併，為每行字幕匹配精準的起訖時間
最終結果：時間軸精準 且 斷句自然 的字幕。
- 精準時間戳 — 基於 ElevenLabs Scribe v2 的字元級對齊
- 語意斷句 — LLM 驅動的智慧換行，尊重語意和閱讀節奏
- 多種字幕風格 — YouTube（完整語句，每行 10–30 字）或 TikTok（短句快節奏，每行 3–8 字）
- 多語言支援 — 中文、英文、日文、韓文、粵語，或自動偵測
- 說話者辨識 — 標註音訊中不同說話者，適合多人對話場景
- Keyterms — 將關鍵詞傳給轉錄 API，提高特定詞彙的辨識準確度
- 文字校正 — 可選的 LLM 校正，修正轉錄錯字
- 翻譯 — 可翻譯為繁體中文、英文、日文或馬來文
- 繁簡轉換 — 使用 OpenCC 輕量轉換，無需呼叫 LLM
- 詞彙表規則 — 定義替換規則（如
OpenAI=>OpenAI）或強制保留詞 - 標點清理 — 移除行首破折號、行尾逗號/句號
- 費用追蹤 — 即時顯示 ElevenLabs 和 OpenAI 的 API 使用成本
- 轉錄模型選擇 — 可選 Scribe v2（推薦）或 Scribe v1
- 多 LLM 模型支援 — GPT-4.1、GPT-4.1 mini、GPT-5.2、GPT-5-mini，可調整推理強度
音訊檔案 (mp3/wav/m4a/flac/ogg)
│
▼
┌─────────────────────┐
│ ElevenLabs Scribe │ → 語音轉錄 + 字元級時間戳
└─────────┬───────────┘
│
▼
┌─────────────────────┐
│ OpenAI LLM │ → 語意斷句，切分為字幕行
└─────────┬───────────┘
│
▼
┌─────────────────────┐
│ 對齊引擎 │ → 為每行字幕匹配精準起訖時間
└─────────┬───────────┘
│
▼
┌─────────────────────┐
│ 後處理 │ → 校正、翻譯、繁簡轉換、標點清理
└─────────┬───────────┘
│
▼
SRT 字幕檔
- Python 3.9+
- ElevenLabs API Key
- OpenAI API Key
git clone https://github.com/leonchanwy/ExactSub.git
cd ExactSub
pip install -r requirements.txtstreamlit run app.py然後開啟終端機顯示的網址（預設：http://localhost:8501）。
- 輸入 API Keys — 在左側欄輸入 ElevenLabs 和 OpenAI 的 API Key
- 調整設定：
- 選擇字幕風格（YouTube / TikTok）
- 設定每行最大字數
- 選擇 OpenAI 模型
- 選擇音訊語言（或自動偵測）
- 上傳音訊檔案（mp3、wav、m4a、flac 或 ogg）
- 點擊「開始生成字幕」
- 下載生成的
.srt字幕檔
| 選項 | 說明 |
|---|---|
| 轉錄模型 | 選擇 Scribe v2（推薦）或 Scribe v1 |
| 說話者辨識 | 標註多人對話中的不同說話者 |
| 自訂斷句指令 | 給 LLM 斷句的額外提示詞 |
| 標點清理 | 移除行首破折號與行尾標點 |
| 字幕校正 | LLM 修正轉錄錯字與聽錯詞 |
| 翻譯 | 透過 LLM 翻譯為其他語言 |
| OpenCC 繁簡轉換 | 輕量簡體→繁體轉換（不需 LLM） |
| 詞彙表規則 | 原詞=>目標詞 替換或保留詞（保留詞同時作為 keyterms 傳給轉錄 API 提高辨識率） |
| 推理強度（GPT-5） | 控制思考深度：none → xhigh |
| 重試設定 | 設定上傳逾時與重試行為 |
# 替換規則（一行一組）
王小明=>小明
OpenAI=>OpenAI
# 保留詞（不含箭頭，維持原樣）
Hokage
應用程式會即時追蹤 API 費用：
| 服務 | 定價 |
|---|---|
| ElevenLabs Scribe | $0.008 / 每分鐘音訊 |
| OpenAI GPT-4.1 | $2.00 / 1M input tokens、$8.00 / 1M output tokens |
| OpenAI GPT-5.2 | $1.75 / 1M input tokens、$14.00 / 1M output tokens |
| OpenAI GPT-5-mini | $0.25 / 1M input tokens、$2.00 / 1M output tokens |