Audio Production & Mastering Skill
A complete pipeline for transcribing, editing, mastering, and bundling spoken audio for Substack podcasts, YouTube replays, community calls, and dev logs.
When to Use
- Transcribing audio files (
.mp3, .wav, .m4a, .webm) into clean Markdown with timestamps and speaker labels.
- Analyzing an unedited recording for dead air, mic feedback, sensitive disclosures, or rambling.
- Performing surgical cuts with smooth crossfades (
acrossfade) so transitions sound completely natural.
- Mastering audio to broadcast/podcast standard (-16 LUFS, 80Hz highpass filter, adaptive noise suppression).
- Packaging deliverables: Always consolidate the cleaned audio and structured markdown transcript into a single unified directory (e.g.
clean_session/ or dist/audio/).
🎙️ Transcription Engines
1. Cloud STT: OpenRouter microsoft/mai-transcribe-2
- Cost/Speed: ~$0.10/hr (half the price of Whisper Large v3 Turbo), 10x faster on long-form audio.
- Key Location:
~/projects/openrouter/.env_openrouter or $OPENROUTER_API_KEY.
- Endpoint:
https://openrouter.ai/api/v1/audio/transcriptions
- 25MB request cap, handled automatically. OpenRouter's own docs cap a
single transcription request at 25MB. A single
transcribe_openrouter()
call over that size will fail outright -- use
transcribe_long_openrouter() for anything that might be long, which
splits by measured bitrate (not a fixed duration guess, since a
32kbps voice memo and a 320kbps studio recording need very different
chunk lengths to land under the same byte cap), transcribes each piece,
and stitches the timestamps back onto the original file's timeline.
Verified 2026-09-15 against a real 57-minute / 110MB / 257kbps file:
558 segments, zero backward-jumping timestamps across chunk boundaries,
last segment's end time matched the file's true duration to five decimal
places, transcribed in 64 seconds for about $0.10.from audio_pipeline import transcribe_long_openrouter
result = transcribe_long_openrouter("recording.m4a") # any length, any size
# result = {"text": "...", "segments": [{"start": 0.0, "end": 7.56, "text": "..."}, ...]}
Or from the CLI: python3 audio_pipeline.py -i recording.m4a --transcribe out.json.
Only reach for the unwrapped transcribe_openrouter() directly when the
input is already known to be under 15MB (its internal safety margin below
the 25MB cap).
2. Local Fallback: faster-whisper
✂️ The Editorial Framework: What to Cut
When reviewing raw streams or community calls, audit for:
- 🚨 Liability & Sensitive Disclosures (Hard Cut if Public):
- Day-job / employer admissions (e.g., coworkers/partners confessing low work hours).
- Regulatory edge cases (Series 65/FINRA compliance disclaimers, unauthorized advice).
- 🔇 Tech Fumbling & Setup Chaos:
- Audio feedback, Discord echo, mic permission loops, teaching someone how to server-mute.
- Start the final cut right where the real substance or opening thesis begins.
- ⏸️ Dead Air & Screen Waiting:
- Multi-second pauses waiting for market open, staring at phones, or frozen screens.
- ⚠️ Self-Deprecating Fluffs & Missing Code:
- Rambling about missing git commits, lost branches, or personal cursing over forgotten UI buttons.
- If producing for a Public Cut, snip cleanly. If producing for a Private Community Cut, keep authentic moments but tighten pauses.
✂️ Doing the Cut: --remove vs --cuts
cut_and_master() (and the --cuts CLI flag) always take KEEP segments --
that's what ffmpeg needs, but it's the wrong shape to reason about when
auditing a transcript for what's bad or too personal. Use --remove
instead: name the ranges you want gone, and invert_segments() computes
the keep-segments around them automatically (sorts unsorted input, handles
removals touching the very start/end, verified with unit tests
2026-09-15). This is the natural fit for the editorial framework above --
read the timestamped transcript, list the [start, end] ranges that match
the Hard Cut / Tech Fumbling / Dead Air / Fluff criteria, pass them to
--remove, done.
python3 audio_pipeline.py -i recording.m4a \
--remove '[[612, 649], [1830, 1905]]' \
-o recording_clean.mp3
🎚️ The Mastering Chain (FFmpeg)
Always apply broadcast mastering to eliminate whisper-to-shout volume swings:
# 1. High-Pass Filter: Strip mic bumps and AC rumble below 80Hz
# 2. Adaptive Noise Reduction: afftdn=nf=-25
# 3. EBU R128 Podcast Loudness: loudnorm=I=-16:TP=-1.5:LRA=11
ffmpeg -i input.mp3 \
-filter:a "highpass=f=80,afftdn=nf=-25,loudnorm=I=-16:TP=-1.5:LRA=11" \
-c:a libmp3lame -b:a 192k output_mastered.mp3
Seam Crossfading
Never hard-cut audio without a transition. Use a 200ms triangular crossfade (acrossfade=d=0.2:c1=tri:c2=tri) between all segments to prevent clicks and room-tone dropouts.
📦 Deliverables Rule: Single Output Folder
Whenever producing cleaned audio and transcripts, never scatter files across root or subdirectories. Always package them together in one dedicated folder:
clean_session/
├── Session_Clean_Public.mp3 # Mastered broadcast audio
└── 2026-09-15_Community_Session_Transcript.md # Formatted Markdown transcript
The Markdown transcript should contain:
- Executive Summary
- Action Items & Dev Bug Log (harvested from stream)
- Full Timestamped Dialogue with Speaker Tags
Default destination for Michael's own recordings
For personal audio (podcasts, voice memos, dev streams -- not a one-off
scratch job), the real, established location is
~/projects/tdpro/mphinance/transcripts/, matching the file already there
(2026-09-15_Community_Session_Transcript.md) -- same naming convention:
YYYY-MM-DD_Descriptive_Name_Transcript.md. This isn't just a transcript
archive: it's raw material for two other things in the same project --
VOICE.md (the writing-style guide, itself mined from "the pre-AI AfterHour
corpus") gets more authentic material to draw from out of how Michael
actually talks, not just what he's already published, and a cleaned
transcript is a legitimate first draft for a Substack post when he asks for
one -- run it past VOICE.md's rules (banned phrases, the sign-off, the tone
rules) rather than treating the raw transcript as publishable as-is.
1---2name: audio-production3description: Transcribe, surgically edit, master, and clean audio recordings (trading streams, community calls, podcasts, meetings) with automated silence/fluff trimming, noise reduction, loudness normalization (-16 LUFS), and transcription via OpenRouter mai-transcribe-2 or local faster-whisper. Always consolidate output into a single dedicated folder.4metadata:5 source: https://skillmd.com/skills/mphinance/audio-production6---78# Audio Production & Mastering Skill910A complete pipeline for transcribing, editing, mastering, and bundling spoken audio for Substack podcasts, YouTube replays, community calls, and dev logs.1112## When to Use1314- Transcribing audio files (`.mp3`, `.wav`, `.m4a`, `.webm`) into clean Markdown with timestamps and speaker labels.15- Analyzing an unedited recording for dead air, mic feedback, sensitive disclosures, or rambling.16- Performing surgical cuts with smooth crossfades (`acrossfade`) so transitions sound completely natural.17- Mastering audio to broadcast/podcast standard (**-16 LUFS**, 80Hz highpass filter, adaptive noise suppression).18- Packaging deliverables: Always consolidate the cleaned audio and structured markdown transcript into a single unified directory (e.g. `clean_session/` or `dist/audio/`).1920---2122## 🎙️ Transcription Engines2324### 1. Cloud STT: OpenRouter `microsoft/mai-transcribe-2`25- **Cost/Speed:** ~$0.10/hr (half the price of Whisper Large v3 Turbo), 10x faster on long-form audio.26- **Key Location:** `~/projects/openrouter/.env_openrouter` or `$OPENROUTER_API_KEY`.27- **Endpoint:** `https://openrouter.ai/api/v1/audio/transcriptions`28- **25MB request cap, handled automatically.** OpenRouter's own docs cap a29 single transcription request at 25MB. A single `transcribe_openrouter()`30 call over that size will fail outright -- **use31 `transcribe_long_openrouter()` for anything that might be long**, which32 splits by measured bitrate (not a fixed duration guess, since a33 32kbps voice memo and a 320kbps studio recording need very different34 chunk lengths to land under the same byte cap), transcribes each piece,35 and stitches the timestamps back onto the *original* file's timeline.36 Verified 2026-09-15 against a real 57-minute / 110MB / 257kbps file:37 558 segments, zero backward-jumping timestamps across chunk boundaries,38 last segment's end time matched the file's true duration to five decimal39 places, transcribed in 64 seconds for about $0.10.40 ```python41 from audio_pipeline import transcribe_long_openrouter42 result = transcribe_long_openrouter("recording.m4a") # any length, any size43 # result = {"text": "...", "segments": [{"start": 0.0, "end": 7.56, "text": "..."}, ...]}44 ```45 Or from the CLI: `python3 audio_pipeline.py -i recording.m4a --transcribe out.json`.46 Only reach for the unwrapped `transcribe_openrouter()` directly when the47 input is already known to be under 15MB (its internal safety margin below48 the 25MB cap).4950### 2. Local Fallback: `faster-whisper`51- Runs directly on CPU/GPU using CTranslate2 int8 quantization, no size cap52 of its own (handles long files directly, no chunking wrapper needed).53 ```python54 from audio_pipeline import transcribe_whisper_local55 result = transcribe_whisper_local("audio.mp3", model_size="base.en")56 # same {"text", "segments"} shape as the OpenRouter path57 ```5859---6061## ✂️ The Editorial Framework: What to Cut6263When reviewing raw streams or community calls, audit for:64651. **🚨 Liability & Sensitive Disclosures (Hard Cut if Public)**:66 - Day-job / employer admissions (e.g., coworkers/partners confessing low work hours).67 - Regulatory edge cases (Series 65/FINRA compliance disclaimers, unauthorized advice).682. **🔇 Tech Fumbling & Setup Chaos**:69 - Audio feedback, Discord echo, mic permission loops, teaching someone how to server-mute.70 - Start the final cut right where the real substance or opening thesis begins.713. **⏸️ Dead Air & Screen Waiting**:72 - Multi-second pauses waiting for market open, staring at phones, or frozen screens.734. **⚠️ Self-Deprecating Fluffs & Missing Code**:74 - Rambling about missing git commits, lost branches, or personal cursing over forgotten UI buttons.75 - If producing for a **Public Cut**, snip cleanly. If producing for a **Private Community Cut**, keep authentic moments but tighten pauses.7677---7879## ✂️ Doing the Cut: `--remove` vs `--cuts`8081`cut_and_master()` (and the `--cuts` CLI flag) always take KEEP segments --82that's what ffmpeg needs, but it's the wrong shape to *reason* about when83auditing a transcript for what's bad or too personal. Use `--remove`84instead: name the ranges you want gone, and `invert_segments()` computes85the keep-segments around them automatically (sorts unsorted input, handles86removals touching the very start/end, verified with unit tests872026-09-15). This is the natural fit for the editorial framework above --88read the timestamped transcript, list the [start, end] ranges that match89the Hard Cut / Tech Fumbling / Dead Air / Fluff criteria, pass them to90`--remove`, done.9192```bash93python3 audio_pipeline.py -i recording.m4a \94 --remove '[[612, 649], [1830, 1905]]' \95 -o recording_clean.mp396```9798---99100## 🎚️ The Mastering Chain (FFmpeg)101102Always apply broadcast mastering to eliminate whisper-to-shout volume swings:103104```bash105# 1. High-Pass Filter: Strip mic bumps and AC rumble below 80Hz106# 2. Adaptive Noise Reduction: afftdn=nf=-25107# 3. EBU R128 Podcast Loudness: loudnorm=I=-16:TP=-1.5:LRA=11108ffmpeg -i input.mp3 \109 -filter:a "highpass=f=80,afftdn=nf=-25,loudnorm=I=-16:TP=-1.5:LRA=11" \110 -c:a libmp3lame -b:a 192k output_mastered.mp3111```112113### Seam Crossfading114Never hard-cut audio without a transition. Use a **200ms triangular crossfade** (`acrossfade=d=0.2:c1=tri:c2=tri`) between all segments to prevent clicks and room-tone dropouts.115116---117118## 📦 Deliverables Rule: Single Output Folder119120Whenever producing cleaned audio and transcripts, **never scatter files across root or subdirectories**. Always package them together in one dedicated folder:121122```text123clean_session/124├── Session_Clean_Public.mp3 # Mastered broadcast audio125└── 2026-09-15_Community_Session_Transcript.md # Formatted Markdown transcript126```127128The Markdown transcript should contain:129- **Executive Summary**130- **Action Items & Dev Bug Log** (harvested from stream)131- **Full Timestamped Dialogue** with Speaker Tags132133### Default destination for Michael's own recordings134135For personal audio (podcasts, voice memos, dev streams -- not a one-off136scratch job), the real, established location is137`~/projects/tdpro/mphinance/transcripts/`, matching the file already there138(`2026-09-15_Community_Session_Transcript.md`) -- same naming convention:139`YYYY-MM-DD_Descriptive_Name_Transcript.md`. This isn't just a transcript140archive: it's raw material for two other things in the same project --141`VOICE.md` (the writing-style guide, itself mined from "the pre-AI AfterHour142corpus") gets more authentic material to draw from out of how Michael143actually *talks*, not just what he's already published, and a cleaned144transcript is a legitimate first draft for a Substack post when he asks for145one -- run it past VOICE.md's rules (banned phrases, the sign-off, the tone146rules) rather than treating the raw transcript as publishable as-is.