DramaBox Cut-Scene Voice-Acting Prompts

Continuously-generated, character-consistent two-scene "CUT TO:" voice-performance prompts (text only, no audio) for training and evaluating expressive TTS / voice-acting models. Each prompt describes a single speaker across two sharply contrasting emotional moments separated by a CUT TO: transition, in DramaBox stage-direction format (spoken lines in "quotes", performance notes in (parentheses)).

• Total prompts: 4,057,000
• Languages: English + German
• Model: unsloth/gemma-3n-E4B-it (bf16, temperature=0.9, max_tokens=1024)
• Avg tokens: input 3251 / output 368
• Shard size: 1,000 rows per parquet file under data/

How it was generated

Prompts are sampled from the Voice-Acting-Pipeline taxonomy across 5 CUT-TO pathways, each producing a character-consistent two-scene prompt with a strong emotional contrast across the CUT TO::

• CCA (VoiceNet) — speaker built from sampled VoiceNet acoustic attributes.
• CC2-C (Archetype) — genre archetype + tempo + arousal.
• ACCC (Acting Challenge) — an acting-challenge brief.
• SIT (Situations) — speaker physically inside a situation that shapes the voice.
• Extreme Physical — calm sensuality erupting into an extreme physical sensation.

Every system prompt carries an emotional-emphasis instruction (make it raw, push the contrast to its limit) and a compact vocal-burst menu (SFW taxonomy, 180 bursts) so the model can place natural non-speech bursts as director notes and use full dynamic range (shout/roar when angry, scream/gasp/sob when afraid or in pain). For ~50% of rows a single vocal burst is randomly injected as a required element (recorded in injected_burst). Dialogue length follows the pipeline default (~50 spoken words, ~25 per scene). Each generation independently samples pathway, language, taxonomy attributes and word-seeds with strong (os.urandom) randomness, so the input state is effectively unique.

• Vocal-burst taxonomy: LAION vocal_bursts_taxonomy_sfw.json (SFW subset of LAION-AI/voice-taxonomies vocalburst) (sfw-v1)
• Burst injection: 2,028,999 with a required burst / 2,028,001 free choice

Distributions

By language

By pathway / condition

By perceived gender

By age group

Top emotions

Top injected vocal bursts

Schema

Each row (one generated prompt) has these columns:

uid, global_index, gen_seed, worker_gpu, pathway, pathway_label, lang, dramabox_prompt, gender, age_group, sampled_gender, sampled_age, arousal, emotions, voicenet_attributes, flow_style, emotion_alignment, direction_style, archetype, genre, tempo, situation, situation_dim, challenge_title, challenge_id, category, subcategory, word_seeds, n_word_seeds, injected_burst, burst_taxonomy_source, burst_taxonomy_version, model, precision, temperature, top_p, max_tokens, input_tokens, output_tokens, gen_timestamp

Key fields: dramabox_prompt (the generated text), pathway/lang, sampled attributes (gender, age_group, arousal, emotions, voicenet_attributes, archetype/genre, situation, challenge_title, category), word_seeds, injected_burst, generation settings (model, temperature, max_tokens), input_tokens/output_tokens, a uid and a global global_index.

License

Released under CC-BY-4.0. Synthetic data generated by a local Gemma model; intended for research on expressive voice acting / TTS. No audio is included.