media-downloader-bot

Audiobook TTS Test

Use bin/zip to exercise the same local-file pipeline used by the bot:

THREADS=1 TTS=OmniVoice /home/braulio/.rvm/wrappers/ruby-3.4.4/ruby bin/zip \
  1908kybalion-pages-1-10.pdf \
  speed=1.0

The default audiobook voice is female, middle-aged, moderate pitch, american accent. Override it with voice= when testing a specific narrator profile:

THREADS=1 TTS=OmniVoice /home/braulio/.rvm/wrappers/ruby-3.4.4/ruby bin/zip \
  1908kybalion-pages-1-10.pdf \
  speed=1.0 \
  voice=male,young_adult,moderate_pitch,american_accent

Non-URL voice= values are passed to OmniVoice as direct voice instructions. Use underscores for spaces inside attributes when calling from the shell, for example young_adult, moderate_pitch, or american_accent.

For audiobook voice cloning, pass an HTTP(S) audio or video link. The audiobook pipeline downloads its best audio stream and reuses the voice-reference quality pipeline to extract the clearest complete passage and transcript:

THREADS=1 TTS=OmniVoice /home/braulio/.rvm/wrappers/ruby-3.4.4/ruby bin/zip \
  1908kybalion-pages-1-10.pdf \
  voice=https://example.com/narrator-recording

transcribe.cpp Subtitles

The TranscribeCpp subtitle backend runs the local CLI and maps its segment and word timestamps directly into the shared Subtitler::Subtitle model:

SUBTITLER=TranscribeCpp \
TRANSCRIBE_CPP_CLI=/path/to/transcribe.cpp/build/bin/transcribe-cli \
TRANSCRIBE_CPP_MODEL=/path/to/canary-1b-v2-timestamps-Q8_0.gguf \
TRANSCRIBE_CPP_LANGUAGE=en \
TRANSCRIBE_CPP_BACKEND=cuda \
TRANSCRIBE_CPP_DEVICE=0 \
bundle exec ruby bin/zip input.wav gensubs onlysrt

The model must advertise word timestamps. Canary requires an explicit language; TRANSCRIBE_CPP_LANGUAGE defaults to en. The adapter converts input media to 16 kHz mono PCM before invoking the CLI. Canary rejects inputs longer than its model limit (about 400 seconds for the current v2 model), so long-form bot media still requires a chunking layer. Canary word timestamps do not include confidence scores and therefore cannot be used by the separate voice-reference quality selector.

Semantic subtitle comparison

Compare two SRT or VTT files by parsed cue text, timing, words, speakers, and VTT presentation metadata:

The reusable API is Subtitler::Subtitle::SemanticDiff.compare(before, after, ...).

bundle exec ruby bin/subtitle_semantic_diff before.srt after.srt \
  --time-tolerance 0.01 --max-details 5

Use --format vtt when the filename extension is unavailable, --json for a machine-readable report, and --fail-on-difference to return status 1 when meaningful differences are found. Input and usage errors return status 2.

Transcription hashtags

Append Codex-generated Instagram-style hashtags to the media caption with any of these options:

bundle exec ruby bin/zip input.wav hashtags lang=pt
bundle exec ruby bin/zip input.wav '#'
bundle exec ruby bin/zip input.wav hts

The generator uses gpt-5.6-luna with low reasoning effort. It follows the requested lang language, chooses singular or plural based on the transcript, and only combines two words when they form a meaningful concept.

Media edits

Use comma-separated time intervals to silence audio or remove sections from audio and video. The same flexible timestamps work for ss=, to=, t=, cuts=, and silences=:

• seconds: 90, 90.5
• clock (M:SS or H:MM:SS), with padding/zeros optional: 1:30, 1:5, :30, 1:, 1::, 1::5, 01:30:00.123
• compact periods: 90s, 1m30s, 1h30m, 1.5m

ss= starts at that offset. to= is an absolute end. t= is a duration after ss= (or from 0); do not combine t= with to=.

bundle exec ruby bin/zip input.mp4 ss=1m30s t=30s
bundle exec ruby bin/zip input.mp4 ss=1:30 to=2:00
bundle exec ruby bin/zip input.mp4 silences=10-20,1:00-1:05
bundle exec ruby bin/zip input.mp4 cuts=30-40,1:30.5-1:45
bundle exec ruby bin/zip input.mp4 cuts=1m-2m,:30-45

silences= preserves the media duration. cuts= removes each interval from both tracks and closes the resulting gaps.

Voice cloning evaluation

Use bin/voice_clone_eval for repeatable OmniVoice clone comparisons. It uses the repository's key=value opts convention. Repeat case= for baseline, reference, and parameter-sweep cases. Each run writes generated audio, transcription scores, speaker-embedding cosine scores, results.json, and summary.csv under a temporary directory and prints the report to stdout:

VOICE_CLONE_EMBEDDING_PYTHON=/srv/sherpa-onnx/runtime/bin/python \
WHISPER_CPP_SERVER=http://127.0.0.1:8080 \
bundle exec ruby bin/voice_clone_eval \
  https://example.com/narrator-recording \
  comparison=source.wav \
  embedding_model=/srv/sherpa-onnx/models/embedding/nemo_en_titanet_small.onnx \
  case=baseline \
  case=raw-default:reference \
  case=guidance-1:reference:guidance=1 \
  case=steps-48:reference:steps=48

The first argument is an HTTP(S) URL. The evaluator downloads it, uses the existing voice-reference transcriber to detect its language, extracts the best voice-reference passage, and uses that passage as the evaluation text and reference text. Results are printed as JSON to stdout; generated artifacts are kept in a temporary directory for the run. Use reference=PATH with text= or text_file= when the reference has already been extracted.

The Sherpa/TitaNet cosine value is a local speaker-similarity proxy, not OmniVoice's official SIM-o metric. Set embedding=false or transcription=false when the corresponding local service is unavailable.

Dubbing timing evaluation

Pass dubscore=PATH to write a JSON timing report for a dubbing run:

bundle exec ruby bin/zip input.mp4 dub=pt dubscore=/tmp/dubbing-timing.json

deviation_index is the root-mean-square absolute log2 tempo adjustment, scaled by 100. A natural-speed run scores 0; both 0.5x and 2x score 100. The report also includes subtitle-slot error and speed distribution values for comparison across future runs.