Base Script CLI Reference

Audio version — Estimated duration: 5 min 53 sec

File: md_to_audio_base.py

Usage

python md_to_audio_base.py <input_md> --ref_audio <wav> --ref_text <transcript> [options]

Arguments

input_md (positional, required)

Path to the input Markdown file. The script reads this file, strips Markdown formatting, and generates speech from the remaining text.

python md_to_audio_base.py sample.md --ref_audio voice.wav --ref_text "..."

If the file doesn’t exist, the script prints an error and exits.

--model_path

Type	Default
str	./Qwen3-TTS-12Hz-0.6B-Base

Path to the Base model directory. This must contain the model files downloaded from Hugging Face (config.json, model safetensors, tokenizer files, etc.).

python md_to_audio_base.py sample.md \
  --model_path /mnt/models/Qwen3-TTS-Base \
  --ref_audio voice.wav --ref_text "..."

--ref_audio (required)

Type	Default
str	— (required)

Path to a WAV file containing the reference voice to clone. The audio should be:

• 5–15 seconds long
• Clean (no background noise, music, or reverb)
• Same speaker throughout
• Sampled at any rate (the model handles resampling)

--ref_text (required)

Type	Default
str	— (required)

The verbatim transcript of --ref_audio. Every word, every filler — “um”, “uh”, pauses — if they’re in the audio, include them. Even small mismatches (“it’s” vs “it is”) noticeably degrade clone quality.

--output

Type	Default
str	output.wav

Path for the output WAV file. The audio is always 24 kHz mono 32-bit float. If the file exists, it will be overwritten.

python md_to_audio_base.py sample.md --ref_audio v.wav --ref_text "..." --output my_clone.wav

--lang

Type	Default
str	English

Language for generation. Must match both the reference audio language and the input text language. Supported values include English, Chinese, Japanese, Korean, French, German, Spanish (depends on the model’s training data).

--chunk_size

Type	Default	Range
int	200	50–500

Maximum characters per chunk. The script splits text at sentence boundaries (.!?), and falls back to commas for long sentences. Smaller chunks use less VRAM but may chop sentences mid-phrase.

Value	Effect
100	Fast, but may sound choppy
200 (default)	Good balance
400	Smoother prosody, more VRAM

--batch_size

Type	Default	Range
int	4	1–16

Number of chunks to process in a single GPU call. Higher values increase throughput but require more VRAM. If a batch fails (OOM), the script falls back to sequential generation automatically.

VRAM	Recommended
6 GB	1–2
8 GB	2–4
12 GB+	4–8

Behaviour details

Model loading

The script loads the model with:

• torch_dtype=torch.float16 — half precision, saves VRAM
• attn_implementation="sdpa" — PyTorch’s scaled dot-product attention
• TF32 matmul and cuDNN kernels enabled (torch.backends.cuda.matmul.allow_tf32, torch.backends.cudnn.allow_tf32)

Generation

Each chunk is generated with max_new_tokens=1024 (fixed). The model produces audio tokens at 12 Hz, so 1024 tokens ≈ 85 seconds of audio — more than enough for any chunk.

Error handling

• If the input Markdown file is empty after cleaning → prints error, exits
• If the model fails to load → prints error, exits
• If a batch fails during generation → falls back to sequential processing per chunk
• If no audio is generated at all → prints error, exits

Output

• Format: WAV
• Sample rate: 24 kHz
• Channels: Mono
• Bit depth: 32-bit float

---

See Voice Cloning Guide for usage examples.