onezion-song-translate

name	onezion-song-translate
description	Chinese → English singable song translation. Translates Chinese lyrics into English while preserving syllable count per line, phoneme repetition patterns (rhyme), musical structure, and semantic meaning — optimized for the lyrics to be sung to the same melody.
agent_created	true

Singable Chinese-to-English Song Translation

Translates Chinese song lyrics into singable English, guided by Peter Low's Pentathlon Principle (2005) and the Computational Evaluation Framework for Singable Lyric Translation (arXiv 2308.13715).

Core Principle: The Pentathlon

A singable translation balances five competing criteria. No single criterion dominates — the goal is a "personal best" across all five:

#	Criterion	Definition	Chinese→English Priority
1	Singability	Physically singable by a vocalist	Open vowels on high notes; avoid consonant clusters at sustained moments
2	Sense	Semantic fidelity to source meaning	Conceptual faithfulness, not word-for-word; key images must survive
3	Naturalness	Sounds like idiomatic English	No translationese; natural register for the genre
4	Rhythm	Syllable count + stress align with melody	Critical: 1 Chinese character ≈ 1 syllable ≈ 1 note; stress must land on strong beats
5	Rhyme	Rhyme scheme preservation	End-rhyme pattern preserved; near-rhyme/assonance acceptable when perfect rhyme would break naturalness

Workflow

Phase 1: Analyze Source (Chinese Lyrics)

For each line of the Chinese lyrics:

1. Count syllables: Chinese = 1 character = 1 syllable (standard Mandarin). Count characters per line.
2. Extract rhyme scheme: Identify end-rhyme pattern (AABB, ABAB, XAXA, etc.) using pinyin finals.
3. Mark musical structure: Label sections (Verse 1, Chorus, Bridge, etc.). Note which sections share melodies.
4. Identify semantic anchors: Underline 2-3 key images/metaphors per section that MUST survive translation.
5. Map stress-to-beat: Mark which syllables land on strong beats (downbeats) — these need English stressed syllables.
6. Flag high notes: Mark lines/syllables sung at high pitch — these need open vowels in English (e.g., /aɪ/, /oʊ/, /ɑː/, not /ɪ/, /iː/).

Output: a structured analysis table.

Phase 2: Translate (Line-by-Line, Constrained)

For each line, apply these constraints in order:

1. Syllable match (strict): Target line must have exactly the same syllable count as source line. Chinese characters → English syllables 1:1 mapping.
2. Stress alignment: English stressed syllables must land on the musical strong beats identified in Phase 1.
3. Vowel openness (high notes): On high/sustained notes, prefer open vowels: /aɪ/ (eye, sky), /oʊ/ (go, flow), /ɑː/ (heart, far), /æ/ (and, dance). Avoid closed vowels: /ɪ/ (it, sit), /iː/ (bee, see), /uː/ (too).
4. End rhyme: Match the target rhyme sound to the rhyme scheme identified in Phase 1.
5. Meaning: Preserve the semantic anchors. Allow line-level reordering if needed for singability.
6. Naturalness: Read the line aloud. If it sounds like "translationese", rephrase.

Phase 3: Evaluate (Four Metrics)

Score the translation using the computational framework:

Metric 1: Syllable Count Distance (Dis_syl)

For each line pair (source i, target i):

Dis_syl = (1/(2n)) * Σ [|syl(X_i) - syl(T_i)| / syl(X_i) + |syl(X_i) - syl(T_i)| / syl(T_i)]

• Target: Dis_syl < 0.05 (near-perfect syllable matching)
• Score 10 = all lines match exactly; Score 1 = avg 50%+ deviation

Metric 2: Phoneme Repetition Similarity (Sim_pho)

For each section, compute phoneme distinct-2 ratio:

pho(section) = unique_bigrams / total_bigrams

(Lower pho = more repetition = more rhyme density)

Then compute Spearman rank correlation between source and target pho vectors across sections.

• Target: Sim_pho > 0.70 (strong correlation)
• Indicates rhyme density patterns are preserved

Metric 3: Semantic Similarity (Sim_sem)

Section-level embedding cosine similarity (using Sentence-BERT or similar):

Sim_sem = Σ [n(X_i)/n_total * cosine_sim(embed(X_i), embed(T_i))]

• Target: Sim_sem > 0.65
• Weighted by section line count

Metric 4: Naturalness Check

Read each line aloud. Score 1-10 based on:

• Does it sound like something a native English speaker would sing?
• No forced word order to achieve rhyme
• Register matches genre (pop → conversational; ballad → slightly elevated)

Phase 4: Iterative Refinement

1. Score Phase 3 metrics
2. Identify the weakest metric
3. Revise targeted lines to improve that metric
4. Re-score
5. Repeat until all metrics meet targets OR 3 iterations completed (pentathlon principle: don't over-optimize one at the expense of others)

Output Format

## 原歌词 / Source Lyrics
[Chinese lyrics with line numbers]

## 分析 / Analysis
| Line | 音节数 | 重拍位置 | 高音标记 | 韵脚 | 
|------|--------|----------|----------|------|
| ...  | ...    | ...      | ...      | ...  |

韵式: [AABB / ABAB / ...]
关键意象: [list of 2-3 per section]

## 英译 / Singable Translation
[English lyrics with line numbers, syllable count annotated]

## 评估 / Evaluation
| Metric | Score | Target | Pass? |
|--------|-------|--------|-------|
| Dis_syl | ...   | <0.05  | ✓/✗   |
| Sim_pho | ...   | >0.70  | ✓/✗   |
| Sim_sem | ...   | >0.65  | ✓/✗   |
| Naturalness | .../10 | ≥7/10 | ✓/✗  |

## 唱法注释 / Singing Notes
- 高音处理: [which words land on high notes, vowel choices]
- 重拍对齐: [stress-beat alignment notes]
- 换气点: [breath/phrasing notes if relevant]

Key References

• Low, Peter. "The Pentathlon Approach to Translating Songs." In Song and Significance: Virtues and Vices of Vocal Translation, ed. Dinda L. Gorlée, Brill, 2005.
• "A Computational Evaluation Framework for Singable Lyric Translation." arXiv:2308.13715 (2023). — Four metrics: Dis_syl, Sim_pho, Dis_mus, Sim_sem.
• Nida, Eugene. Toward a Science of Translating. — Principle of "fixed length per musical phrase, identical syllable count."
• "Sing it, Narrate it: Quality Musical Lyrics Translation." EMNLP 2024 Findings. — Constrained translation with length and rhyme requirements.

Chinese→English Specific Notes

• 1 character = 1 syllable: Unlike English where words are multi-syllable, each Chinese character produces exactly one syllable. This makes syllable counting deterministic — no syllabification algorithm needed.
• Tone → Stress proxy: Mandarin tones (1-4) don't map directly to English stress, but tone 1 (high level) and tone 4 (falling) often align with musical downbeats and can be treated as stress anchors.
• Pinyin finals for rhyme: Rhyme in Chinese is carried by the pinyin final (e.g., -ang, -ai, -an, -i, -u). Group finals into rhyme classes for pattern detection.
• Open vowel substitution: When a Chinese closed vowel (i, ü, -i) falls on a high note, the English translation should use an open vowel even if the semantic word has a closed vowel — use a synonym or rephrase.