natural-sounding-narration-with-frankentts
Made by JeffDirect and repair FrankenTTS narration with stable voice references, contextual takes, deliberate pauses, exact-sample edits, transcription checks and human listening. Use when generating ftts voice-over or fixing accent drift, cadence, pronunciation or pacing.
From Jeffrey's Skills.md: professionally crafted skills for any AI agent, installed and kept in sync by the jsm CLI. Subscription only, $20/month.
Works With
Quality and Utility come from canonical skill metadata. Community ratings appear separately as stars when subscriber reviews exist.
Subscribe to download
Public previews are available, but downloads, CLI sync, and premium content require an active subscription.
Sign In to Subscribe- Full skill content with deterministic versioning
- CLI sync + local search index
- Access to Jeffrey’s premium library
Save to your account
Saved skills live in your cloud list. Run jsm sync on each machine to pull them locally.
What's New with this Skill
2Updates, enhancements, and release notes for this skill.
Updated: natural-sounding-narration, catches the clipped and dropped words that ASR lets through
Three narrated explainers passed every gate, with an ASR match near 0.99, and their owner rejected all three by ear: noise after short sentences, sentences that stopped inside their last word, quiet consonants cut out, and "a million phone" for "phones". Whisper had written the right words every time. The skill now reads takes as spectrograms, reads each sentence again in quotation marks so ftts finishes it, cuts only at real silence, and flags a missing final s or z.
- Each sentence is also read in quotation marks, so ftts finishes its last word
- The editor measures silence against each take's own speech level and stops cutting quiet consonants
- A missing or clipped final s or z is caught from the spectrum, which the ASR can't hear
- Spectrogram sheets of every take and line are read before the gate
Found while rebuilding the narration of four math explainers. The new chooser was then run on all four finished narrations, and every pick and line came out the same except one: a read the old check had marked down by mistake.
- Quoted reads. ftts stops about half of plain sentences inside their last word; in one video, 158 of 241 chosen reads were still within 25 dB of speech level 50 ms before the file ended. Each sentence is also read inside quotation marks, which ftts finishes. The quotes go to the TTS only, so the script and captions don't change.
New skill: natural-sounding-narration-with-frankentts, voice-actor narration from FrankenTTS with every name looked up and verified
An agent turns a script into FrankenTTS narration that sounds like a voice actor who knows the material. It looks up how every name and term is pronounced, turns the IPA into spellings FrankenTTS reads correctly, and checks each with FrankenWhisper first. Each sentence gets several takes; FrankenWhisper scores them, and the best are joined with pauses that follow the meaning and mastered with one gain. Speech is never time-stretched, and nothing ships until QA and a stretch check pass.
- Every name and term is looked up: lookup_pronunciation.py reads IPA from the Wikipedia article's lead and from Wiktionary, falls back to a low-confidence guess from the spelling, and flags conflicting variants and acronyms that can be read as letters or as a word.
- ipa_to_respelling.py turns the IPA into spellings FrankenTTS reads correctly, such as Koh-shee for Cauchy and foor-YAY for Fourier. Respellings change only the text sent to TTS; the script and captions keep the real spelling.
- probe_names.py says each respelling at the start, middle and end of a sentence, plus a control (the ordinary spelling) and a decoy (a wrong reading), and FrankenWhisper judges the result by sound. Passing spellings go into lexicon.json with their IPA and sources.
- Controls and decoys catch ASR bias: two different readings of Euler both came back as "Euler", so Euler is marked unverified instead of passing. In a 12-term test, 8 terms passed in every position and 4 were flagged for a human listen.
- The owner's pronunciation overrides every source. After that comes a person's own usage, then the pronunciation in the name's own language, with the English variant shown beside it.
- Each sentence is its own FrankenTTS call, and each take has different text: greedy decoding makes a new seed return identical audio, so takes come from respellings and director alternates.
- FrankenWhisper transcribes every take. The score weights the match 0.40, confidence and pace 0.15 each, and pauses, pitch movement and contour 0.10 each, minus 0.25 per misheard name. Statements must fall at the end and yes/no questions rise.
- Speech is never time-stretched or pitch-shifted. Pace comes from wording, take choice and pause length; slowing sentences by up to 15% once put echo and phasey distortion on exactly those sentences.
- Only cuts and gain touch the audio: pauses capped at 0.36 s, hesitations cut to 0.14 s, sentences matched to −23 LUFS, a limiter at −8 dBFS, and one fixed-gain master toward −16 LUFS under a −1.5 dBTP ceiling, with no loudnorm.
- Nothing ships until qa_lines.py passes (mean FrankenWhisper match of 0.97 or better, no line under 0.90, no misheard name) and validate_narration.py confirms no line was stretched. On a published explainer's 61 lines, the QA mean was 0.999.
FrankenTTS reads text only, with no IPA or SSML input, so pronunciation is controlled by the spelling sent to it; the script and captions keep the real spelling. lookup_pronunciation.py finds how any name or term is pronounced in public sources, with no API key: the IPA in the lead of its Wikipedia article (English usage and the person's own language), then Wiktionary, and only when neither has IPA a rule-based guess from the spelling, marked low confidence. It flags conflicting variants and acronyms that can be read as letters or as a word. ipa_to_respelling.py turns the IPA into two or t