Script2Voice

Generate narration WAV files and matching SRT subtitles from a tagged script.

tagged script -> Qwen3-TTS longform WAV -> Qwen3 ForcedAligner -> SRT

Each [tag] block becomes one narration module. Script2Voice exports one WAV/SRT pair per module and also writes a stitched audio_full.wav with SRT_FULL.srt.

Input

Tags and visual notes must be on their own lines:

[summary]
(Show project homepage)
大家好，今天介绍一个本地 AI 视频知识管理工具。

[settings]
第一次使用时，进入左侧的控制中心进行配置。

[summary] starts a module. (Show project homepage) is saved as a visual note in blocks.json, but it is not sent to TTS and does not appear in subtitles.

Configuration

Machine-level settings live in settings.toml. These settings usually stay the same across projects:

[tts]
model = "models/Qwen3-TTS-12Hz-1.7B-Base"
device = "cuda"
dtype = "bf16"
max_seq_len = 8192
max_new_tokens = 4096

[aligner]
model = "Qwen/Qwen3-ForcedAligner-0.6B"
device_map = "cuda:0"
dtype = "bf16"
max_chunk_seconds = 90.0

[paths]
hf_hub_cache = "cache/hf"

Use models/ for models you manage manually, such as a local Qwen3-TTS folder. Use cache/ for files downloaded automatically by HuggingFace or model libraries. max_chunk_seconds limits each forced-alignment pass so long narration audio does not overflow GPU memory.

Voice and style presets live in presets/. A preset is what you choose per video style:

[voice]
ref_audio = "voices/tutorial/ref.wav"
ref_text_file = "voices/tutorial/ref.txt"

[style]
instruct = "请用自然、清晰、稳定的中文教程讲解语气朗读。"
temperature = 0.6
top_k = 50
repetition_penalty = 1.1

Put the reference WAV and its exact transcript in the matching voices/<name>/ folder. Real voice files and transcripts are ignored by git.

Quick Start

1. Put the TTS model under models/, or edit settings.toml to point to your existing model path.

models/
  Qwen3-TTS-12Hz-1.7B-Base/

2. Put your voice reference audio and transcript under voices/.

voices/
  tutorial/
    ref.wav
    ref.txt

ref.txt must contain the exact words spoken in ref.wav.

3. Configure a preset in presets/tutorial.toml.

[voice]
ref_audio = "voices/tutorial/ref.wav"
ref_text_file = "voices/tutorial/ref.txt"

[style]
instruct = "请用自然、清晰、稳定的中文教程讲解语气朗读。"
temperature = 0.6
top_k = 50
repetition_penalty = 1.1
greedy = false
xvec_only = true
non_streaming_mode = false

[continuity]
enabled = true
ref_tail_seconds = 30.0
min_ref_seconds = 8.0

Create more presets when you want different voices or delivery styles:

presets/
  tutorial.toml
  product_intro.toml
  calm_narration.toml

When continuity is enabled, the first tag uses the preset reference voice. Later tags use the previous generated block's tail as a temporary reference, which can improve cross-block voice and mood consistency. ref_tail_seconds caps the reference length, and min_ref_seconds avoids using references that are too short.

4. Write a tagged script.

[summary]
(Show project homepage)
大家好，今天介绍一个本地 AI 视频知识管理工具。

[settings]
第一次使用时，进入左侧的控制中心进行配置。

5. Generate WAV and SRT.

E:\gittools\self\tagged-tts-blocks\tagged-tts.cmd `
  --script "E:\video_process\videos\5月16日\vsummary_script.txt" `
  --output-dir "E:\video_process\videos\5月16日\vsummary_voice" `
  --preset tutorial

To ignore tag boundaries and generate one full-length narration block:

E:\gittools\self\tagged-tts-blocks\tagged-tts.cmd `
  --script "E:\video_process\videos\5月16日\vsummary_script.txt" `
  --output-dir "E:\video_process\videos\5月16日\vsummary_full_voice" `
  --preset tutorial `
  --full

Full mode sets max_new_tokens to 25600 for that run and disables rolling reference continuity because there is only one generated block.

Output

Output:

output/
  audio_blocks/
    001_summary.wav
    002_settings.wav
  subtitle_blocks/
    001_summary.srt
    002_settings.srt
  audio_full.wav
  SRT_FULL.srt
  blocks.json

Per-block SRT files start at 00:00:00,000. SRT_FULL.srt is offset to match audio_full.wav.

Dependencies

The wrapper uses:

E:\conda-envs\ai-cu128\python.exe

Required Python packages in that environment:

python -m pip install qwen-tts qwen-asr soundfile numpy

Models are not stored in this repository. By default, place them under models/ or edit settings.toml.

Notes

Script2Voice does not generate TTS sentence by sentence. It generates each tag as longform audio first, then uses forced alignment to derive sentence-level subtitle timing.