Wenbi

Wenbi is a CLI-first toolkit that turns media (video/audio/URL) and text into structured Markdown, then rewrites or translates the result. It is built around the ollama/qwen3.5:cloud rewrite/translation model by default, with DeepL as the preferred translator and an LLM fallback.

It supports:

• Video/audio/URL transcription to VTT/Markdown
• Text rewriting (rewrite, academic style)
• English interview rewriting (en-en) with speaker-separated output, now bilingual by default (English → Chinese)
• Chinese interview rewriting (zh-zh) with speaker-separated output
• English/Chinese bilingual audio extraction (en-zh) — keep English, translate to Chinese
• Single-language multi-speaker diarization (speaker) with rewrite + translate
• Optional slide-combine (--ppt / -p on any subcommand) — align slides with speech by timestamp
• Batch directory processing (wenbi-batch)

How it works

Every subcommand follows the same pipeline; the only difference is which stages run and the default options:

input (media / URL / text / subtitles)
  │
  1. Download & extract        ── yt-dlp for URLs, ffmpeg/pydub for media → audio
  │
  2. ASR / transcribe          ── auto | gladia | funasr | whisper
  │                                gladia is the default when GLADIA_API_KEY is set;
  │                                funasr (paraformer-zh, local) is the fallback;
  │                                whisper is the third fallback.
  │                                → VTT + raw transcript
  │
  3. Diarize (optional)        ── gladia / funasr cam++ → speaker labels
  │
  4. Rewrite (optional)        ── LLM (default ollama/qwen3.5:cloud)
  │                                rewrites oral speech into written style,
  │                                preserves speaker turns for interview flows
  │
  5. Translate (optional)      ── DeepL first (needs DEEPL_API_KEY),
  │                                LLM fallback if DeepL unavailable/fails.
  │                                EN→ZH glossary applied by default
  │                                (DeepL glossary API + LLM prompt).
  │
  6. Slide combine (--ppt only)  ── frame extraction + optional OCR → slides aligned
  │                                with speech by timestamp.
  │                                Bare --ppt: embed frames as base64 (TYPE 1, no OCR).
  │                                --ppt <file>: OCR the slides file, match to frames (TYPE 3).
  │                                Bilingual subcommands: timestamps recovered from VTT
  │                                via rapidfuzz fuzzy matching before alignment.
  │
  └── outputs in --output-dir:   *_rewritten.md, *_translated.md,
                                  *_bilingual.md, *_zh.md, *_en.md,
                                  *_combine.md, *_diagnostics.json, *.vtt, *.csv

Subcommand → pipeline mapping:

Install

Prerequisites:

• Python 3.11+
• ffmpeg in PATH
• Gladia API key (optional but recommended): set GLADIA_API_KEY for cloud ASR (the default for all audio-capable subcommands). Free tier is 10h/month — see https://www.gladia.io/pricing. Without the key, ASR auto-falls-back to FunASR (local, no key needed, requires the heavy ML deps).

Install:

pip install wenbi

Or from a local checkout:

# from the project directory
pip install -e .

Quick Start

Rewrite:

wenbi rewrite input.mp4 --lang Chinese --llm ollama/qwen3.5:cloud

English interview rewrite:

wenbi en-en interview.mp4 --gladia-key "$GLADIA_API_KEY"

Chinese interview rewrite:

wenbi zh-zh interview.mp4 --gladia-key "$GLADIA_API_KEY"

English/Chinese bilingual audio (keep English, translate to Chinese):

wenbi en-zh bilingual.mp4 --gladia-key "$GLADIA_API_KEY"

Single-language multi-speaker (diarize, rewrite, translate):

wenbi speaker panel.mp4 --source-lang en --gladia-key "$GLADIA_API_KEY"

Slide-combine (any subcommand + --ppt):

# TYPE 1: embed video frames as base64, combine with rewritten speech
wenbi rewrite lecture.mp4 --ppt

# TYPE 3: OCR an external slides file, match to video frames, combine
wenbi rewrite lecture.mp4 --ppt slides.pdf

Commands

rewrite (rw)

Rewrite spoken/transcribed text into written style.

wenbi rewrite <input> [options]

Key options:

• --style rewrite|academic
• --lang
• --llm
• --asr-provider auto|gladia|funasr|whisper
• --cite-timestamps
• --start-time, --end-time (media/URL)

en-en (enen)

Transcribe an English interview, separate speaker turns, rewrite it as polished written English, and translate to Chinese by default (DeepL first, LLM fallback). Runs stages 1 → 2 → 3 → 4 → 5.

wenbi en-en <input> [options]

Key options:

• --speaker-count (default: 2)
• --asr-provider auto|gladia|funasr|whisper
• --gladia-key (or GLADIA_API_KEY env var)
• --llm (default: ollama/qwen3.5:cloud)
• --lang — target translation language (default: Chinese; set to English to skip translation and keep English-only output)
• --deepl-key (or DEEPL_API_KEY env var)
• --glossary / --no-glossary (default: enabled, EN→ZH term consistency)
• --glossary-file path.json (custom {english: chinese} glossary)
• --start-time, --end-time (media/URL)

The rewrite preserves speaker labels and adds a ## Questions for Clarification section when speaker roles, names, terms, or ambiguous ASR phrases need human confirmation.

zh-zh (zhzh)

Transcribe a Chinese interview, separate speaker turns, and rewrite it as polished written Chinese using ollama/qwen3.5:cloud by default.

wenbi zh-zh <input> [options]

Key options:

• --speaker-count (default: 2)
• --asr-provider auto|gladia|funasr|whisper
• --gladia-key (or GLADIA_API_KEY env var)
• --llm (default: ollama/qwen3.5:cloud)
• --start-time, --end-time (media/URL)

The rewrite preserves speaker labels and adds a ## 需要确认的问题 section when speaker roles, names, terms, or ambiguous ASR phrases need human confirmation.

en-zh (enzh)

Extract English from English/Chinese bilingual audio (e.g. interpreted interviews), drop the interpreter language, and translate the kept English into Chinese using DeepL first with LLM fallback.

wenbi en-zh <input> [options]

Key options:

• --asr-provider auto|gladia|funasr|whisper (default: auto)
• --source-lang (default: en) — language to keep
• --interpreter-lang (default: zh) — language to drop
• --gladia-key (or GLADIA_API_KEY env var)
• --lang — target translation language (default: Chinese)
• --glossary / --no-glossary (default: enabled, EN→ZH term consistency)
• --glossary-file path.json (custom {english: chinese} glossary)
• --no-speaker-labels — disable speaker diarization
• --save-json — write segment diagnostics and raw provider JSON
• --start-time, --end-time (media/URL)

Outputs include the kept-language VTT/Markdown, a bilingual Markdown side-by-side, and (optionally) a rewritten English Markdown and diagnostics JSON.

speaker (sp)

Transcribe single-language multi-speaker audio with diarization, then rewrite and translate it. Same engine as en-en/zh-zh but without the interview-style rewrite defaults — use it for panels, podcasts, and any multi-speaker source where you want to keep the source language.

wenbi speaker <input> [options]

Key options:

• --asr-provider auto|gladia|funasr|whisper (default: auto)
• --source-lang (default: en)
• --speaker-count (default: provider decides)
• --gladia-key (or GLADIA_API_KEY env var)
• --lang — target translation language (default: Chinese)
• --glossary / --no-glossary (default: enabled, EN→ZH term consistency)
• --glossary-file path.json (custom {english: chinese} glossary)
• --no-speaker-labels — disable speaker diarization
• --save-json — write segment diagnostics and raw provider JSON
• --start-time, --end-time (media/URL)

Outputs a transcript VTT, transcript Markdown, rewritten Markdown, and (when translation is requested) a bilingual Markdown.

Slide-Combine (--ppt / -p)

Any subcommand accepts --ppt / -p to add stage 6 (slide-combine) on top of its own stages. Useful for lectures and talks where slides should be aligned with the transcript.

wenbi <subcommand> <input> --ppt [slides_file] [options]

Two modes:

• Bare --ppt / -p (TYPE 1): extract video frames, deduplicate, embed as base64 images with timestamps. No OCR — frames are the slides.
• --ppt <file> / -p <file> (TYPE 3): OCR the given PPT/PDF/image file, match pages to video frames via SSIM, combine. Higher quality when a clean slides file is available.

Slide options (apply to both modes):

• --frame-interval — extract a frame every N seconds (default: 60)
• --max-slides, --no-deduplicate, --dedup-method, --similarity-threshold, --ssim-threshold, --hist-threshold
• --no-clean — keep timestamps and image references in combined output
• --no-ocr — TYPE 3 only: embed slide images as base64 instead of OCR'ing them

Outputs {base}_combine.md (and {base}_combine_clean.md unless --no-clean).

Timestamp alignment

Slide-combine aligns slides to speech by timestamp (### **HH:MM:SS** headers). Two paths:

• rewrite subcommand: produces timestamped speech directly (forces --cite-timestamps when --ppt is set). Slides align precisely.
• Bilingual subcommands (en-zh, en-en, zh-zh, speaker): the rewritten output strips timestamps (clean prose separated by ---). --ppt recovers timestamps by fuzzy-matching each rewritten paragraph against the VTT file from stage 2 ASR using rapidfuzz. Since group_into_topics preserves 100% of transcript text and rewrite_english keeps ~97% wording, the match scores high and the correct VTT segment window is found for each paragraph. The recovered timestamps are then used for slide alignment.

Supported Inputs

• Media: .mp4 .avi .mov .mkv .flv .wmv .m4v .webm .mp3 .flac .aac .ogg .m4a .opus
• Text/subtitles: .vtt .srt .ass .ssa .sub .smi .txt .md .markdown .docx
• URL inputs are supported for media flows.

Common Global Options

Used by subcommands:

• --output-dir
• --lang
• --llm
• --translation-engine auto|deepl|ollama|openai — openai selects openai/gpt-5.6-terra and disables DeepL
• --chunk-length
• --max-tokens
• --timeout
• --temperature
• --asr-provider auto|gladia|funasr|whisper
• --transcribe-lang
• --multi-language
• --verbose

OpenAI GPT-5.6-terra translation

Use the OpenAI Platform API key as an environment variable; do not put it in a command, config file, or source code. A ChatGPT/Codex subscription does not itself supply an API key or API credits.

export OPENAI_API_KEY="sk-..."
uv run wenbi speaker "https://www.youtube.com/watch?v=tFX_IVRAcXI&t=18183s" \
  --translation-engine openai --lang Chinese --asr-provider gladia -v

With --translation-engine openai, the topic grouping, English cleanup, and Chinese translation all use GPT-5.6-terra; DeepL and Ollama are not used. GPT-5 models only accept temperature=1, so Wenbi applies that value even when the general CLI default is 0.1.

Glossary

Translation-capable subcommands (en-en, en-zh, speaker) apply a built-in EN→ZH patristic glossary by default for term consistency. This covers ~800 Orthodox Christian / patristic terms (e.g. theosis → 神化, theoria → 静观, Origen → 奥利金).

• --glossary / --no-glossary — toggle the glossary (default: enabled). No-op when the target language is not Chinese.
• --glossary-file path.json — supply a custom glossary JSON ({english: chinese} dict). Overrides the built-in patristic glossary.

The glossary is wired into both translation paths:

• DeepL: creates a DeepL server-side glossary from the term pairs and passes it to translate_text.
• LLM fallback: injects the glossary as a glossary field on the DSPy TranslateSignature prompt.

Output Files

Typical outputs:

• *_rewritten.md
• *_translated.md
• *_academic.md
• *_en.md, *_en.vtt (English interview transcripts)
• *_zh.md, *_zh.vtt (Chinese interview transcripts)
• *_bilingual.md (en-zh and speaker translated output)
• *_diagnostics.json (when --save-json is used)
• *_combine.md / *_combine_clean.md (PPT workflows)
• *.vtt, *.csv (depending on flow)

Batch Processing

Process a directory of media files:

wenbi-batch <input_dir> --output-dir <dir> --md

Optional config:

wenbi-batch <input_dir> --config config.yaml

YAML Config (CLI)

wenbi supports YAML via --config.

Example:

input: lecture.mp4
output_dir: ./out
llm: ollama/qwen3.5:cloud
lang: Chinese
chunk_length: 20

Multi-input format is also supported using inputs:.

Python API

from wenbi.main import process_input

text, md_file, csv_file, base_name = process_input(
    file_path="input.mp4",
    subcommand="rewrite",
    lang="Chinese",
    llm="ollama/qwen3.5:cloud",
)

Troubleshooting

• No DeepL translation output:
  • Set DEEPL_API_KEY or --deepl-key
  • Run with --verbose to confirm DeepL connectivity
• Fallback LLM not working:
  • Ensure your provider is reachable (for example, Ollama running locally for ollama/...)
• PPT OCR issues:
  • Ensure marker_single and OCR dependencies are installed correctly

License

Apache-2.0