#7080 what is new in 4.12

2a09**db Posted at: 1 hour ago

what is new in 4.12

#1 DeepSeek reply at 1 hour ago

当前知识库覆盖至 v4.11，不包含 v4.12 的更新详情，无法提供准确的变更内容。

建议：

前往官方发布页面查看版本说明：https://github.com/jianchang512/pyvideotrans/releases
或直接升级到 v4.11 后，检查软件内是否有更新提示（如果 v4.12 是未公开版本，请以官方正式发布为准）。

若你已安装 v4.12 并遇到具体问题，请提供软件左上角版本号、操作系统及报错日志。

#2 2a09**db reply at 59 minutes ago

in the video there is hardcoded subtitles that is in there at the start, how can i make it so the dubbed audio match the original subtitles, and what is the best asr settings to match the orignal video

#3 DeepSeek reply at 59 minutes ago

pyVideoTrans cannot extract hardcoded (burned-in) subtitles – it only recognizes the audio track. To make the dubbed audio match the original subtitles, you must first extract those subtitles to an SRT file and import them as the source.

Solution: Use external SRT as the source

Extract the hardcoded subtitles with a separate tool (e.g. video-subtitle-extractor or OCR tools). Get an .srt file.
Place it in the correct folder structure (default output location):
- For video myvideo.mp4, create a subfolder _video_out/myvideo-mp4/ next to the video.
- Put the SRT inside, named with the source language code, e.g. en.srt.
- If you want the translated subtitles too, put zh-cn.srt there as well.
In the main interface: select the video, DO NOT check "清理已生成". The software will then skip ASR/translation and directly use your SRT for dubbing and merging.
Enable sync alignment in the alignment row:
- Check "配音加速" (audio acceleration) – if the dubbing is longer than the original subtitle timing, it will be sped up.
- Optionally check "视频慢速" – if you prefer to slow the video instead.
- If you want the final embedded subtitles to match the dubbed audio precisely, enable "二次识别" (but note this only works with non-hardcoded embedded subtitle modes).

Best ASR settings if you still use automatic recognition

If you cannot extract the subtitles and must rely on ASR, no settings can guarantee an exact match with hardcoded subtitles (ASR transcribes speech, not burned-in text). However, for better alignment with the video's natural speech:

Channel: faster-whisper(内置) with large-v3-turbo model
Advanced Options → Speech Recognition:
- 最长语音持续(秒): 6–10
- 最短语音持续(毫秒): 3000–4000
- Enable 合并过短字幕到邻近
- Enable VAD (default is fine)

Please consult related documents:

#4 2a09**db reply at 57 minutes ago

answer in english

#5 DeepSeek reply at 57 minutes ago

pyVideoTrans cannot extract hardcoded (burned-in) subtitles – it only recognizes the audio track. To make the dubbed audio match the original subtitles, you must first extract those subtitles to an SRT file and import them as the source.

Solution: Use external SRT as the source

Extract the hardcoded subtitles with a separate tool (e.g. video-subtitle-extractor or OCR tools). Get an .srt file.
Place it in the correct folder structure (default output location):
- For video myvideo.mp4, create a subfolder _video_out/myvideo-mp4/ next to the video.
- Put the SRT inside, named with the source language code, e.g. en.srt.
- If you want the translated subtitles too, put zh-cn.srt there as well.
In the main interface: select the video, DO NOT check "清理已生成". The software will then skip ASR/translation and directly use your SRT for dubbing and merging.
Enable sync alignment in the alignment row:
- Check 配音加速 (audio acceleration) – if the dubbing is longer than the original subtitle timing, it will be sped up.
- Optionally check 视频慢速 – if you prefer to slow the video instead.
- If you want the final embedded subtitles to match the dubbed audio precisely, enable 二次识别 (but note this only works with non-hardcoded embedded subtitle modes).

Best ASR settings if you still use automatic recognition
If you cannot extract the subtitles and must rely on ASR, no settings can guarantee an exact match with hardcoded subtitles (ASR transcribes speech, not burned-in text). However, for better alignment with the video's natural speech:

Channel: faster-whisper(内置) with large-v3-turbo model
Advanced Options → Speech Recognition:
- 最长语音持续(秒): 6–10
- 最短语音持续(毫秒): 3000–4000
- Enable 合并过短字幕到邻近
- Enable VAD (default is fine)

请查阅相关文档：

#6 2a09**db reply at 56 minutes ago

how about other settings in asr

#7 DeepSeek reply at 56 minutes ago

Here are the remaining ASR settings you can tune in 菜单 → 工具 → 高级选项 → 语音识别参数. The most important ones for aligning dubbing with the original video’s speech timing are already covered, but these additional parameters let you refine accuracy, reduce hallucination, and control resource usage.

Setting	What it does	Recommended value for matching original timing
`选择VAD`	Selects the Voice Activity Detection model used to split speech segments.	Keep default unless you have a specific need.
`语音阈值`	Minimum probability for a segment to be considered speech. Lower = more sensitive but may pick up noise.	Default is fine. If you get extra silent lines, increase slightly.
`非语音阈值`	Threshold for marking a segment as non-speech. Lower reduces hallucination but may drop real speech.	Default. If you see repeated empty lines, increase a bit.
`静音分割持续毫秒`	The silence duration required before splitting a new speech segment.	Default ~500ms. If subtitles are too fragmented, increase to 800–1000ms.
`合并过短字幕到邻近`	Merges subtitles shorter than the `最短语音持续` value into adjacent lines.	Always enable. This reduces tiny useless subtitle lines.
`Whisper预分割音频`	Pre-splits audio into sentence-like chunks before sending to Whisper. This can improve timing accuracy.	Enable if you use `clone` voice. Set `最短语音持续` to 3000ms and `最长语音持续` to 10s.
`计算数据类型`	`int8` = low memory, faster, lower accuracy; `float32` = highest accuracy, more memory; `float16` = best for GPU.	Use `float16` if you have an NVIDIA GPU with >6GB VRAM, else `int8`.
`beam_size` / `best_of`	Decoding search depth. Higher = more accurate but consumes more VRAM and time.	Keep at `1` for speed. Set to `5` only if you need maximum accuracy and have VRAM.
`启用上下文感知`	Uses previous text context to improve recognition.	Disable for timing alignment – it can cause repetition or hallucinations.
`重复惩罚`	Penalizes repeated text.	If you get loops, increase to `1.2–1.5`.
`文本压缩率`	Compression ratio for decoding.	Lower values reduce repetition. Try `0.9` if repeats appear.
`采样温度`	Sampling temperature. Higher = more randomness.	Keep default.
`热词`	List of terms you want the model to recognize correctly (comma‑separated).	Add domain‑specific words for better accuracy, e.g. `AI, deep learning`.
`字幕繁体转简体`	Converts traditional Chinese subtitles to simplified.	Enable if your original video is in Chinese traditional.
`删除字幕末尾标点`	Removes trailing punctuation from each subtitle line.	Not needed for timing.
`云API识别暂停秒`	Pause between cloud API requests to avoid rate‑limit errors.	Only relevant for cloud ASR channels.
`二次识别最长/最短语音`	Same as above, but applied during the second recognition pass after dubbing.	Only relevant if you enable `二次识别`. Use shorter values (e.g. max 6s, min 500ms) to get short, tightly‑aligned subtitles.

Key reminder: No ASR settings can make the recognized subtitles exactly match the hardcoded burned‑in subtitles – ASR transcribes the audio, not the burned‑in text. If you already have the original SRT, use the external‑SRT method described earlier — that is the only reliable way to achieve frame‑perfect subtitle‑to‑audio alignment.

#7080 what is new in 4.12

7 Replies

Solution: Use external SRT as the source

Best ASR settings if you still use automatic recognition

Please consult related documents:

请查阅相关文档：

请查阅相关文档：

Post Your Reply

Related resource