JP Text to JP-EN Audio Pipeline

Turns a Japanese source text (e.g., a novel) into listening-first study material: sentence-aligned Japanese/English pairs, plus optional short “micro-breakdowns” (up to three per sentence). Some sentences may also include an optional “literal” English line to make the Japanese structure easier to see.

The pipeline: clean + split the raw text into manageable chunks → translate sentence-by-sentence with batch context (optionally adding micro-breakdowns + a literal line) → generate per-sentence audio files you can play like an “audiobook” with JP/EN pairs (plus optional breakdown/literal audio).

Scripts

divide.py — clean + split raw text into API-sized batches

• Input: jp.txt
• Output: divided_output.txt
• Cleans the text first by removing ruby/furigana and transcription markup (e.g., 《...》, ｜, ［＃...］).
• Splits the text into paragraphs (blank-line separated), then groups paragraphs into batches capped by MAX_CHARS_PER_BATCH (and MAX_PARAGRAPHS_PER_BATCH).
• If a paragraph is too long, it is split at Japanese sentence boundaries first, then by line breaks, then by character-count as a fallback.
• Batches in the output are separated by --- (specifically \n\n---\n\n).

deepseek_breakdown.py — translate + optional micro-breakdowns (DeepSeek API)

• Input: divided_output.txt
• Output: translated_output.json (+ incremental translation_progress.json)
• For each batch: splits into sentences and numbers them, then sends them together for contextual translation. Sentence splitting tries not to break inside Japanese quotes like 「...」 / 『...』.
• Expects a JSON array with fields: sentence_number, english, optional english_literal, and up to 3 optional breakdown pairs: part_to_breakdown_1..3 and breakdown_1..3.
• Micro-breakdowns are meant to be short: each one pairs a JP phrase (part_to_breakdown_*) with a brief English explanation (breakdown_*).
• Saves the final output as a dict with metadata plus a sentences object keyed by global sentence index (each entry includes japanese, english, optional english_literal, batch_number, and breakdown fields).
• Reads API settings from config.json (or environment variables), with basic retry/backoff.

gen_audio.py — synthesize audio for JP/EN pairs (Balabolka)

• Input: translated_output_part.json
• Output folder: audio_sentences/
• Uses Balabolka bal4web.exe with Microsoft voices; auto-detects working JA male/female voices and tries to find an EN voice.
• Generates audio per sentence as MP3s (WAVs are generated as intermediates and then converted/cleaned up).
• Generates multiple segments per sentence (where possible): sentence number (JP), Japanese sentence (male), English translation, combined breakdown audio (up to 3 JP terms + EN notes), optional literal-English audio, and an alternating JP female→male version.
• Uses configurable pauses/silence and per-segment speech-rate settings; retries transient TTS failures.

Example files included

• jp.txt — example input Japanese text (No Longer Human / 人間失格 by Osamu Dazai)
• divided_output.txt — example output from divide.py
• translated_output.json — example output from deepseek_breakdown.py

Notes

• This repo is not actively maintained.
• Upload the scripts to an AI assistant if you want a deeper, line-by-line explanation.

License

GPL.