qwen3-tts-italian-expr

qwen3-tts expressivity assets (emotion + paralinguistics)

Composable expressivity add-ons for qwen3-tts (the pure-C inference engine for Qwen3-TTS). These are plugins layered on top of normal synthesis — they do not replace the base model. All for the 1.7B model.

What's here — organized by purpose

You normally don't touch these files directly — --emotion <name> and inline [tags] load the right ones automatically (see How to activate). The raw --expr/--ml-steer flags are an advanced manual override.

The tiny steer/ vectors also ship inside the GitHub repo (presets/steer/). The big expr/ files live here on HF because they are too large for git.

expr/ — emotion fine-tunes

• italian_csp_topk6.expr — Italian emotion default (cleanest on presets + clones).
• german_csp_k6.expr, french_csp_k6.expr — native German / French emotion.
• italian_csp_topk4.expr, italian_l1626_dense.expr (+ _r32/_r64, multi/multitag) — earlier Italian variants (A/B / research).

steer/emotion/ — ryan_{ang,sad,joy,fear,disgust,surprise}, galatea_{ang,sad}_ft, vivian_{ang,sad}_ft

steer/paraling/ — laugh_vs_cry, sigh_vs_laugh (+ .qamp source captures to rebuild)

How to download

# from the qwen3-tts repo:
bash download_assets.sh            # fetches expr/ into presets/expr/ (sha256-verified)
# or grab a single file:
curl -L -o presets/expr/italian_csp_topk6.expr \
  https://huggingface.co/gabrione/qwen3-tts-italian-expr/resolve/main/expr/italian_csp_topk6.expr

Disk: the full expr/ set ≈ 1.4 GB; Italian-only emotion needs just italian_csp_topk6.expr (203 MB).

How to activate — it's automatic

You don't wire these files by hand. The engine composes the right stack for you.

Emotion → one flag, --emotion. Pick an emotion and the engine applies the ear-validated recipe: a preset voice → pure STEER (the steering vector ryan_<emo> @ w12 — clean, every language); a cloned voice → COMBINE (the language .expr + steer). Use the native preset per language (JA ono_anna, KO sohee, ZH vivian, EN/Romance ryan). A vivid English --instruct is an optional override.

# emotion in ONE flag — engine picks expr + steer automatically
./qwen_tts -d qwen3-tts-1.7b -s ryan -l Italian -T 1.1 --emotion sad \
  --instruct "Speak softly, with quiet sadness." \
  --text "Allora, lascia che ti spieghi come stanno le cose." -o sad.wav

--emotion accepts: sad · joy · anger · fear · disgust · surprise (synonyms like happy/angry work too).

Works in every Qwen3-TTS language, not just IT/DE/FR. The steering vector carries the emotion (it's a language-agnostic activation direction); the .expr renders/stabilizes the language. DE / FR / IT have a native .expr; for ES / PT / RU / ZH / JA / KO the engine uses the Italian pack as a universal cross-language renderer (ear-validated on IT/ES/RU/ZH/JA/KO). Just set -l <Language> and go.

Paralinguistics → inline [tags] in the text. No flags: write [laugh] or [sigh] in --text and the engine performs the event automatically (it picks the onomatopoeia anchor + the right vector for you).

./qwen_tts -d qwen3-tts-1.7b -s ryan -l Italian -T 1.1 \
  --text "Che giornata... [sigh] non ce la faccio più. [laugh]" -o para.wav

Advanced / manual override. Power users can still wire the raw pieces — --expr <file> --expr-weight, --ml-steer <file> --ml-range 21-25 --ml-weight <w> — and a manual flag always overrides the auto-router. Full recipes, per-voice weights and the tuning notes: docs/expressivity-assets.md.

License / attribution

CC-BY 4.0. The Italian emotion fine-tune uses the Emozionalmente dataset — please cite: F. Catania, J. W. Wilke, F. Garzotto (Politecnico di Milano), IEEE TASLP 33:1142-1155, 2025, doi:10.1109/TASLPRO.2025.3540662.