Elysia Tools
Navigation
Workflow Playbook
Map where speech lives in a finished recording, transcribe it into captions in the format the player expects, align caption timing to the speech, mix in audio description with program ducking, and prove the message survives degraded listening.
New to this workflow? Follow the steps in order, starting with the first tool below. Already know what you need? Jump straight to the matching step — every tool in this workflow is listed in the comparison table.
First tool: Audio Voice Activity Segmentation
Hubs
| Tool | What it does | Category |
|---|---|---|
| Audio Voice Activity Segmentation | Detect likely speech-range activity and export speech/non-speech ranges as JSON, CSV, SRT markers, or an FFmpeg cut list. | Media |
| Audio to Text Transcriber (AI) | Transcribe speech from audio (wav/mp3/m4a/flac/ogg/webm/aac) to text, SRT, VTT or JSON using the grok-stt AI model. Up to 10 minutes. | AI Tools |
| Audio Caption Timing Shifter | Shift SRT or WebVTT captions to a likely speech onset or by a manual offset. | Media |
| Audio Description Ducking | Mix an audio-description narration into program audio while ducking the program when narration is present. | Media |
Hubs
| Assistive Listening Enhancer | Clarify listening audio with low-frequency cleanup, speech-presence EQ, conservative denoising, and level protection. | Media |
| Emergency Alert Intelligibility Simulator | Test whether an alert retains heuristic speech clarity after bandwidth, compression, and noise degradation. | Media |
A finished spoken-audio cut serves listeners with quiet rooms, sharp hearing, and full attention. Everyone else negotiates: deaf and hard-of-hearing listeners need the words as text, blind listeners need the visual world narrated, hearing-aid users need a track that protects levels while clarifying speech, and a public alert must survive phone speakers and street noise. Accessibility work is not charity padding after production — it is a second delivery surface with its own artifacts, formats, and verification. This workflow produces that surface from one finished program: a speech map, captions in the format the destination expects, description narration mixed over ducking, and intelligibility that holds under the worst realistic listening.
Everything downstream hangs on knowing where speech actually lives. Voice-activity segmentation walks the finished program and returns speech and non-speech ranges as JSON, CSV, SRT markers, or an FFmpeg cut list. The map does three jobs at once — it bounds what must be captioned, it reveals the non-speech windows where description narration can sit without fighting dialogue, and it exposes dead air that a listening track should not force anyone through. A caption floating over silence is a defect the map would have caught, which is why the map comes first.
Transcription turns the speech into text, and the format decision belongs to the destination — SRT or VTT for players with caption tracks, plain text for documents and review, JSON for pipelines that process captions further. Then comes the step that separates usable captions from embarrassing ones — timing alignment. Uniformly late or early captions shift to the likely speech onset; a known constant delay, like a broadcast chain or a player buffer, takes a measured manual offset. Progressive drift is a different animal — it means the caption file and the media disagree about time itself, usually a frame-rate mismatch, and the honest fix is re-deriving captions against the correct source. Verification is cheap and non-negotiable — first caption, last caption, one in the middle.
Audio description narrates the visual world for listeners who cannot see it, and it rides over a program that ducks out of the way while the narration speaks. The craft is in the two settings — depth and release. Depth is set by the loudest passage that runs underneath narration; if the words read clearly there, they read everywhere. Release is set by the program's own rhythm — music and dialogue should bounce back promptly instead of sinking into a hole. The failure mode to check is the single worst case, not the average — one unintelligible narration entry makes the described version skip, exactly where a blind listener cannot fill the gap from the picture.
The last step answers the question production rarely asks — does the message still work when listening is hard? For the hearing-access track, conservative enhancement cleans the low end, biases the EQ toward speech presence, denoises gently, and above all protects levels, because a clarified track that clips has traded one barrier for another. For public alerts, the degradation test is stricter on purpose — it narrows the bandwidth to a phone speaker, compresses, adds noise, and reports whether the alert's speech clarity survives each channel. A warning that only works on studio monitors has not been tested; it has been hoped at.
This page produces the accessibility layer of one finished program. The neighbors hold the surrounding jobs. Recording, editing, and delivery compliance for an episode belong to the podcast production workflow. Measuring, visualizing, and checking — spectrograms, waveform, caption-sync checks — belong to the audio analysis pages. A source recording that is damaged — clicks, hum, wind — belongs in restoration repair before any accessibility work starts, because captions of uncorrected audio inherit its defects. Caption the program here — and when the audience grows, the discipline scales with it.
Workflow playbook
Run voice-activity segmentation over the finished program to get speech and non-speech ranges as JSON, CSV, SRT markers, or a cut list — the map tells you what needs captions and where description narration can sit.
Convert the speech to text with output as SRT, VTT, plain text, or JSON, choosing the format the destination expects rather than accepting a default.
Shift captions to the likely speech onset for uniform misalignment, or apply a measured manual offset when a known constant delay exists, then verify at the start, middle, and end of the program.
Blend the description narration over the program so the program ducks while narration speaks and recovers when it stops, keeping every narration entry intelligible at the loudest underlying passage.
SRT and VTT suit video players with caption tracks, plain text suits documents and review copies, and JSON suits automated pipelines — choose by where the captions are consumed, not by what the transcriber happened to emit.
Shifting to the likely speech onset fixes captions that sit uniformly late or early; a measured manual offset fixes a known constant delay, such as a broadcast chain or player buffer — drifting timing across the program means re-checking the source, not nudging once.
The program must drop far enough for the narration to read clearly, yet release fast enough not to punch a hole in music or dialogue — depth serves the narration, release time serves the program.
For the hearing-access track, apply conservative enhancement — low-frequency cleanup, speech-presence EQ, level protection; for public alerts, run the degradation test that reports whether speech clarity survives phone speakers, compression, and noise.