Preset Voices (CustomVoice Model)

Audio version — Estimated duration: 5 min 28 sec

Use one of 9 built-in voices with no reference audio required. The CustomVoice model (Qwen3-TTS-12Hz-0.6B-CustomVoice) has speaker embeddings baked in — just pick a name.

Available speakers

Name	Likely language / style	Notes
Aiden	English, neutral	Default speaker
Serena	English, warm	Good for narratives
Vivian	English, bright
Uncle_Fu	Chinese / accented English	Distinctive character voice
Ryan	English, deep	Good for announcements
Ono_Anna	Japanese / English	Bilingual capable
Sohee	Korean / English	Bilingual capable
Eric	English, clear	Good for tutorials
Dylan	English, casual

All speaker names are case-insensitive — the script resolves "serena" → "Serena" automatically. If a name doesn’t match, it falls back to "serena" and prints the available list.

Basic usage

python md_to_audio_custom.py sample.md --speaker "Vivian"

Style instructions

The --instruct parameter controls delivery — think of it as a stage direction for the speaker:

python md_to_audio_custom.py sample.md \
  --speaker "Ryan" \
  --instruct "Speak slowly and clearly, like a news anchor."

Instruction examples:

Instruction	Effect
Speak naturally. (default)	Neutral, conversational
Speak slowly and clearly.	Slower pace, crisper enunciation
Read with enthusiasm and energy.	Higher pitch variation, faster
Whisper quietly.	Softer delivery
Speak in a sad and melancholic tone.	Lowered pitch, slower
Read like a dramatic movie trailer.	Exaggerated emphasis, pauses

The instruction is passed to every chunk, so keep it short (1 sentence). Longer instructions may degrade quality.

Sampling parameters

These are hardcoded in the script and control the randomness of generation:

Parameter	Value	Effect
temperature	0.8	Lower = more consistent, higher = more varied
top_k	50	Only consider the 50 most likely tokens at each step
top_p	0.95	Nucleus sampling — cumulative probability cutoff

For most use cases the defaults are fine. If you want to modify them, edit the call to model.generate_custom_voice() in md_to_audio_custom.py around line 175.

Chunk size and speed

The script defaults to --max-chars 150 — tighter than the Base model’s 200. This is intentional:

--max-chars	Trade-off
100	Fastest per-chunk, ~15 audio tokens generated, but may chop sentences
150 (default)	Good balance for 6 GB VRAM
200–250	Smoother prosody, more VRAM per chunk
300+	May overshoot VRAM on 6 GB cards with batch_size > 1

Adaptive token budget

The CustomVoice script doesn’t use a fixed max_new_tokens like the Base script. Instead, estimate_max_tokens() calculates a per-chunk budget:

text_tokens ≈ len(chunk) / 4.5 chars per token
audio_tokens ≈ text_tokens × 12 Hz × 1.3 safety margin
clamped: max(256, min(audio_tokens, 1024))

This means a 20-character chunk gets max_new_tokens=256 instead of 1024 — roughly 4× faster for that chunk. The speedup is most noticeable on short documents (< 1000 characters).

Full examples

Article with character voice:

python md_to_audio_custom.py story.md \
  --speaker "Uncle_Fu" \
  --instruct "Tell this folk tale warmly, like a grandfather." \
  --max-chars 180 \
  --output story.wav

Tutorial with clear instruction:

python md_to_audio_custom.py tutorial.md \
  --speaker "Eric" \
  --instruct "Speak clearly and precisely, emphasizing key terms." \
  --dtype bfloat16 \
  --output tutorial.wav

---

Next: CLI Reference · GPU Optimization