Bazarr Comparison

⚖️ RightSub vs. Bazarr — Technical & Code-Level Architectural Comparison

Language / שפה: English | עברית

---

📌 Executive Summary & Architectural Conclusions

This technical audit is based on Bazarr's official documentation (Bazarr Subtitles Wiki) and a line-by-line inspection of Bazarr's open-source GitHub codebase (morpheus65535/bazarr).

The Core Question: "Why do I need RightSub if Bazarr is already running on my media server?"

Short Answer: Because Bazarr is an Ingestion Web Crawler & Downloader, while RightSub is a Semitic Linguistics, BiDi Engineering, and Mastering Factory.

Users frequently observe buttons in Bazarr's UI labeled Reverse RTL, Remove HI Tags, OCR Fixes, Translate, and Two-Point Fit, and assume Bazarr provides a complete subtitle solution for Hebrew. Inspecting the underlying source code reveals that these tools were inherited from the legacy Sub-Zero library, engineered specifically for Western European Latin languages. When executed on Semitic (Hebrew/Arabic) subtitles, they fail across four fundamental dimensions:

1. Physical BiDi Corruption: Bazarr's Reverse RTL performs a crude physical character swap instead of injecting logical Unicode directionality marks, corrupting subtitles on modern players (Apple TV, Infuse, Plex, Android TV).
2. Total Blindness to Hebrew in Sanitization: Bazarr's OCR and Hearing Impaired filters contain zero Hebrew Unicode codepoints (\u0590-\u05FF) in their regular expressions and dictionaries, silently skipping Hebrew files.
3. Flat, Gender-Blind Machine Translation: Translation tools execute a static 3-line prompt on massive 300-cue batches with zero pre-flight analysis, zero TMDb cast gender resolution, and zero context overlap, causing severe second-person grammatical gender flips ("אתה" vs. "את") and severed sentences.
4. Torrent Seed Invalidation: In-place file edits inside media folders alter file sizes and cryptographic hashes, causing active torrent clients (qBittorrent, Transmission) to fail with I/O hash check errors.

Engineering Verdict: RightSub and Bazarr are not rivals; they are complementary. The optimal automated home media setup uses Bazarr to crawl and download community files from the web, and triggers RightSub as a Post-Processing hook to master, sanitize, and format the subtitles to broadcast standards.

---

🏛️ Code-Level Head-to-Head Analysis

1. Directionality & BiDi Punctuation: Reverse RTL vs. SubRefine

• In Bazarr Source Code (custom_libs/subzero/modification/mods/common.py, lines 119–134):

  class ReverseRTL(SubtitleModification):
      identifier = "reverse_rtl"
      processors = [
          NReProcessor(re.compile(r"(?u)(^([\s.!?:,'-]*)(.+?)(\s*)(-?\s*)$)"), r"\5\4\3\2", name="CM_RTL_reverse")
      ]

• The Engineering Flaw:
  • The routine assumes the target playback engine lacks Unicode BiDi support and performs a crude physical swap of line characters (\5\4\3\2).
  • Modern media players (Plex, Apple TV, Infuse, Android TV, VLC) implement native Unicode Bidirectional Algorithm (UAX #9).
  • When Bazarr's physical swap is applied, dialogue dashes jump from the start of the line to the end (שלום- instead of - שלום). Furthermore, the regex trailing capture only matches hyphens (-?\s*$), completely ignoring trailing question marks (?) and exclamation points.
• The RightSub Solution (SubRefine):
  • Keeps 100% of the underlying text in valid, logical UTF-8 sequence. Never physically reverses characters.
  • Injects invisible Unicode Right-to-Left Marks (\u200F - RLM) adjacent to neutral punctuation and dialogue dashes in strict compliance with Unicode UAX #9.
  • Result: Flawless rendering on Apple TV, Plex, Infuse, and Smart TVs without character inversions.

---

2. Auditory Noise Sanitization: Remove HI Tags vs. RightSub SDH Cleaner

• In Bazarr Source Code (custom_libs/subzero/modification/mods/hearing_impaired.py, lines 60–64):

  NReProcessor(re.compile(r'(?sux)-?%(t)s[\"\']*\[(?=[^\[\]]{3,})[A-Za-zÀ-ž0-9\s\'\".:-_&+]+[\)\]][\"\'*[\s:]*%(t)s' % {"t": TAG}), "", name="HI_brackets")

• The Engineering Flaw:
  • The bracket regex character class [A-Za-zÀ-ž0-9...] strictly matches Latin and Western European characters.
  • Hebrew Unicode codepoints (\u0590-\u05FF) are completely excluded. Tags such as [מוזיקה מתנגנת], (טלפון מצלצל), or [צעקות] are never matched and remain intact.
  • Additionally, the All-Caps sound description list is hardcoded to exactly 19 English-only words (LAUGH, SCREAM, DOOR, etc.).
• The RightSub Solution (junk_and_sdh_cleaning.py):
  • Full native support for Hebrew and English auditory patterns across all bracket types ([], (), {}).
  • Removes musical notation (♪) while strictly preserving cue timing intervals.
  • Whitelist protection (test_legitimate_dialogue_not_flagged_as_ad) to ensure short dialogue lines are never mistakenly stripped.

---

3. OCR and Typographical Repairs: OCR Fixes vs. Homoglyphs & Semantic Polish

• In Bazarr Source Code (custom_libs/subzero/modification/mods/ocr_fixes.py):

  data_dict = OCR_fix_data.get(parent.language.alpha3t)
  if not data_dict:
      logger.debug("No SnR-data available for language %s", parent.language)
      return

• The Engineering Flaw:
  • Inspecting dictionaries/data.py reveals dictionary tables for only 16 European languages:
    ['bos', 'dan', 'deu', 'eng', 'fin', 'fra', 'hrv', 'hun', 'nld', 'nob', 'nor', 'por', 'rus', 'spa', 'srp', 'swe'].
  • Hebrew (heb) is totally unsupported. When clicked on a Hebrew subtitle, Bazarr outputs a debug log entry and silently exits without modifying anything.
• The RightSub Solution (v1.3.0):
  • Universal Homoglyph Normalization: Automatically identifies and replaces foreign character intrusions introduced by legacy bitmap OCR (e.g. Cyrillic с resembling Latin c or Hebrew characters).
  • Hebrew Typographical Normalization: Converts ASCII double quotes into canonical Hebrew Gershayim (״ / \u05F4) for acronyms (e.g., עו״ד, ארה״ב, ד״ר).
  • Semantic AI Polish (rightsub polish): Uses LLMs and 3-tier domain lore (domain_knowledge.py) to audit spelling, slang, and franchise canon.

---

4. Legacy Encodings & Mojibake: Raw Pass-Through vs. Auto-Charset Engine

• In Bazarr:
  • Bazarr crawls files from community websites (OpenSubtitles, Subscene, Torec) and saves them with their original encoding.
  • Many older Israeli subtitle archives were saved in Windows-1255 (CP1255) or ISO-8859-8.
  • When modern streaming players read CP1255 files under UTF-8 assumptions, subtitles display as unreadable gibberish (àðé øåöä instead of אני רוצה). Bazarr has no automated Semitic encoding detection or conversion pipeline.
• The RightSub Solution:
  • The Auto-Charset engine evaluates all input files with statistical heuristic confidence, automatically detecting CP1255 and ISO-8859-8 encodings and converting them to clean, modern UTF-8.

---

5. Timing Synchronization: Two-Point Fit vs. Headless CLI & Hardware-Accelerated STT

• In Bazarr Source Code (custom_libs/subzero/modification/mods/two_point_fit.py):

  parent.f.shift(...)
  parent.f.transform_framerate(float(kwargs.get("from")), float(kwargs.get("to")))
  parent.f.shift(...)

• The Operational Difference:
  • The underlying linear time scaling and offset math matches RightSub's timing_and_sync.py.
  • The Bazarr Bottleneck: Requires manual user interaction: opening the video in an external player, seeking out two spoken lines, recording 4 separate timestamps, typing them into a web GUI modal, and clicking save.
• The RightSub Solution:
  • Single-Line Headless CLI: FPS conversions and millisecond offsets execute in a single headless command (rightsub sync).
  • Hardware-Accelerated Alignment (quicksubs): Performs local on-device speech-to-text audio alignment against the video's audio track via Apple Silicon Neural Engine/Metal or local Whisper models — zero manual timestamp hunting and zero cloud API expenses.

---

6. Translation & Linguistics: Bazarr's Flat Engine vs. RightSub's Pre-Flight Architecture

This is the primary architectural differentiator refuting the misconception that Bazarr's "Translate" button makes an external tool redundant.

How Bazarr Executes Translation (bazarr/subtitles/tools/translate/services/):

• Google Translator (google_translator.py): Dispatches individual cues in parallel via ThreadPoolExecutor(max_workers=10). Every line is translated in complete isolation: zero conversational context, total grammatical gender blindness, and broken sentence flow.
• Gemini Translator (gemini_translator.py, lines 135–150): Sends monolithic 300-cue blocks using a static, 3-line prompt:

  You are an assistant that translates subtitles to {language}.
  Dialogs must be translated as they are without any changes.
  If a line has a comma or multiple sentences, try to keep one line to about 40-50 characters.
  

  • Zero Pre-Flight Intelligence: No TMDb character cast query, no speaker extraction, and no awareness of speaker identities or genders.
  • Zero Context Overlap: If a dialogue sentence is split across subtitle batches, the first half is translated in batch A and the second half in batch B without conversational context.
  • Destructive BiDi Encodings: Bazarr wraps Hebrew lines in legacy Unicode 2.0 \u202b...\u202c (RLE/PDF) directional embedding codes. In modern media players (Apple TV, Infuse, Plex), these obsolete marks physically invert numbers, times, and currency symbols.

RightSub's 3-Phase Pre-Flight Architecture:

1. Phase 1: Pre-Flight Bible & TMDb Entity Resolution (03_generate_bible.py):
   • Before submitting a single token to the LLM, RightSub scans the English SRT for explicit speaker tags (ALAN:, [DENNY]), character names, and recurring honorifics (Judge, Counselor, Dr.).
   • Executes an automated pre-flight query against the TMDb API to resolve the verified cast roster, character names, and official genders (Male/Female).
   • Generates and locks translation_bible.json, guaranteeing that character gender and nomenclature remain 100% consistent across entire 24-episode seasons.
2. Phase 2: Master System Prompt & Context Overlap (09_prompt_builder.py):
   • Explicit Vocative Agreement Rules: Enforces strict Semitic 2nd-person gender inflection ("אתה" vs. "את", "אתה רוצה" vs. "את רוצה") based on speaker and addressee continuity.
   • SDH Filtered as Context-Only: Speaker tags like [Tara] or ALAN: are supplied as context to inform the LLM of conversational participants, while the prompt strictly forbids outputting these metadata tags into final subtitles.
   • Sliding Context Overlap: Every batch receives 3–5 overlapping dialogue lines from preceding and succeeding cues, preventing fragmented sentences.
   • Domain Knowledge & Canon Lore (domain_knowledge.py): Dynamically injects genre glossaries (legal courtroom terms, medical jargon, sci-fi franchises).
3. Phase 3: Semantic AI Polish (19_polish_and_qc.py):
   • If Bazarr already downloaded an existing community subtitle that suffers from machine-translation errors, RightSub does not need to re-translate the file from scratch.
   • The rightsub polish engine operates on Minimal Edit Distance, auditing the Hebrew subtitle against master English dialogue to preserve 85%–90% of valid human phrasing while surgically fixing gender flips and terminology at ~15% of the token cost.

---

7. Torrent Seeding Protection: In-Place Mutation vs. Seed-Safe Architecture

• In Bazarr:
  • When Bazarr edits, converts, or synchronizes a subtitle file inside a media folder, it alters the file in-place.
  • This mutates the file size and payload content. BitTorrent clients (qBittorrent, Transmission) detect hash mismatches during re-checks and immediately halt torrent seeding with I/O errors.
• The RightSub Solution:
  • Operates on a Strictly Non-Destructive Seed-Safe Model by default.
  • Leaves the original downloaded subtitle 100% bit-for-bit intact, creating an external <stem>.he.srt that Plex and Infuse automatically recognize and prioritize.

---

8. Resource Footprint & System Impact: 24/7 Daemon vs. Zero-Daemon CLI

• Bazarr:
  • Designed as an always-on server daemon or continuous Docker container running 24/7.
  • Occupies port 6767, maintains background worker loops, and consumes continuous RAM (150–300 MB).
• RightSub:
  • Zero-Daemon Architecture.
  • Lightweight CLI utility with zero background processes.
  • Idle resource consumption: 0% CPU and 0 MB RAM. Wakes up on event triggers or manual calls, masters the subtitle in ~0.2s, and immediately terminates.

---

📊 Comprehensive Comparison Matrix (Version 1.3.0)

Capability / Dimension	Bazarr	RightSub (v1.3.0)
Primary Pipeline Role	Automated web crawler & downloader for existing community subtitles (Ingest).	Mastering, Semitic linguistics, BiDi RLM, AI proofreading, and headless automation (Mastering).
Plex & Infuse BiDi Handling	❌ Broken physical character swap (common.py) or legacy RLE tags (\u202b).	✔️ SubRefine engine injecting invisible Unicode RLM (\u200F) marks (UAX #9).
Hearing Impaired / SDH Sanitization	❌ Latin characters only; blind to Hebrew auditory descriptions.	✔️ Native bilingual SDH sanitizer with dialogue protection.
OCR Typo Repair	❌ Hebrew is completely missing from dictionary (data.py).	✔️ Homoglyph normalization from 7 scripts + Hebrew Gershayim normalization.
Legacy Mojibake Repair (CP1255)	⚠️ Passes through raw encoding from provider.	✔️ Auto-Charset engine detecting Windows-1255 and converting to UTF-8.
Timing Synchronization (Sync)	⚠️ Manual web GUI only (Two-Point Fit requires typing 4 timestamps).	✔️ Direct CLI flags (rightsub sync) + hardware-accelerated STT (quicksubs).
Pre-Flight Architecture & Bible	❌ None. Static 4-line prompt with zero domain knowledge.	✔️ TMDb cast resolution, speaker extraction, locked translation_bible.json.
Context Overlap Across Batches	❌ Flat 300-line batches. Sentences severed without context.	✔️ 3–5 cue sliding overlap windows guaranteeing sentence continuity.
Vocative & Gender Agreement	❌ Completely blind (defaults to masculine, frequent gender flips).	✔️ Explicit 2nd-person gender rules based on speaker continuity.
Subtitle Proofreading / Polish	❌ None. Requires translating from scratch.	✔️ rightsub polish with Minimal Edit Distance (85%–90% preservation).
Torrent Seeding Protection (Seed-Safe)	❌ Modifies files in-place, corrupting torrent hashes.	✔️ Non-destructive duplication to <stem>.he.srt (Bit-for-Bit intact).
NAS / SMB Metadata Synchronization	None (susceptible to macOS SMB 2001 epoch bug).	Explicit os.utime() metadata sync for immediate NAS indexing.
Automated 24/7 Web Crawling	✔️ Industry standard (crawls 30+ providers linked to Sonarr/Radarr).	❌ Does not crawl pirate sites; processes local files or generates subtitles via AI.
System Resource Overhead	24/7 Docker container / background daemon (RAM/CPU).	Zero-daemon CLI running purely on event hooks (0% idle CPU, 0 MB RAM).

---

🤝 The Recommended Synergy: Bazarr + RightSub

The most reliable media server architecture pairs Bazarr for ingestion with RightSub for mastering:

In Bazarr (Settings -> Subtitles -> Post-processing -> Custom Post-Processing):

rightsub auto "{{subtitles_path}}"

The End-to-End Workflow:

1. Acquisition: Bazarr monitors media libraries and downloads raw community subtitles from web providers.
2. Autonomous Mastering (0.2s): RightSub triggers immediately upon download, converts CP1255 to clean UTF-8, injects invisible RLM marks for Plex/Apple TV, removes SDH noise and ad spam, preserves torrent seeding, and updates NAS timestamps.
3. AI Fallback & Polish: If Bazarr finds no subtitle on the web, or retrieves a low-grade machine translation, RightSub can translate or polish the media file directly at broadcast quality.