Voice & Audio Production

name: voice-audio-production
version: 1.0.0
updated: 2026-02-21
author: InstaClaw
triggers:
  keywords: [voiceover, voice, audio, narration, TTS, text to speech, podcast, speech, sound, voice message, narrator]
  phrases: ["add voiceover", "make a voiceover", "generate audio", "text to speech", "podcast intro", "voice message", "narrate this", "read this aloud", "audio version", "add narration to video"]
  NOT: [play music, music production, voice call, phone call, transcribe audio, speech to text]

Overview

Generate professional voiceovers, audio content, and voice messages using text-to-speech APIs. The killer use case: Remotion videos with synchronized voiceovers — transforming silent motion graphics into broadcast-quality content.

Prerequisites (already on your VM):

• FFmpeg (format conversion, mixing, normalization)
• Remotion (video+audio integration)
• OpenAI API key (TTS endpoint — same key as LLM, always available)
• ElevenLabs API key (premium voices — check .env for ELEVENLABS_API_KEY)

Provider Selection

Use ElevenLabs when:

• Public-facing content (marketing videos, podcasts, demos)
• Voice quality is critical
• User is on Pro or Power tier

Use OpenAI TTS when:

• Internal/draft content
• Voice messages via Telegram
• Document summaries
• User is on Free/Starter tier
• ElevenLabs is unavailable or over monthly limit

Always check before generating:

1. Run python3 ~/scripts/audio-usage-tracker.py check <char_count> to verify budget
2. If ElevenLabs is over limit, fall back to OpenAI TTS automatically
3. Log usage after every generation: python3 ~/scripts/audio-usage-tracker.py track <chars> <provider>

Workflow 1: Remotion Video with Voiceover (THE KILLER FEATURE)

This is the end-to-end pipeline from "make me a video" to final MP4 with synchronized voiceover. Generate audio FIRST, then set video duration to match. Never the other way around.

Step 1: Write the Voiceover Script

// Write the script based on the video's purpose
const script = `Welcome to InstaClaw. The AI agent platform that works 24/7.
Traditional chatbots just answer questions. InstaClaw agents take action.
Deploy your own AI agent in 60 seconds. Try it today.`;

// Estimate duration (~150 words per minute for natural speech)
const wordCount = script.split(/\s+/).length;
const estimatedSeconds = (wordCount / 150) * 60;
console.log(`Script: ${wordCount} words, ~${estimatedSeconds.toFixed(1)}s`);

Step 2: Generate Audio

Option A: ElevenLabs (premium)

# Using the helper script
~/scripts/tts-elevenlabs.sh "$SCRIPT" public/voiceover.mp3

# Or directly via API
curl -s -X POST "https://api.elevenlabs.io/v1/text-to-speech/21m00Tcm4TlvDq8ikWAM" \
  -H "xi-api-key: $ELEVENLABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"text\": \"$SCRIPT\", \"model_id\": \"eleven_monolingual_v1\", \"voice_settings\": {\"stability\": 0.5, \"similarity_boost\": 0.75}}" \
  --output public/voiceover.mp3

Option B: OpenAI TTS (standard/fallback)

# Using the helper script
~/scripts/tts-openai.sh "$SCRIPT" public/voiceover.mp3 alloy tts-1-hd

# Or directly via API
curl -s -X POST "https://api.openai.com/v1/audio/speech" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"model\": \"tts-1-hd\", \"voice\": \"alloy\", \"input\": \"$SCRIPT\"}" \
  --output public/voiceover.mp3

Step 3: Get Exact Audio Duration

# CRITICAL: Use ffprobe to get exact duration — don't estimate
DURATION=$(ffprobe -v error -show_entries format=duration -of csv=p=0 public/voiceover.mp3)
echo "Audio duration: ${DURATION}s"

Step 4: Remove Leading/Trailing Silence

# TTS engines often add 0.5-1s of silence. Remove it for tight sync.
~/scripts/audio-toolkit.sh silence-remove public/voiceover.mp3 public/voiceover-clean.mp3
mv public/voiceover-clean.mp3 public/voiceover.mp3

# Re-measure duration after silence removal
DURATION=$(~/scripts/audio-toolkit.sh duration public/voiceover.mp3)

Step 5: Normalize Volume

~/scripts/audio-toolkit.sh normalize public/voiceover.mp3 public/voiceover.mp3

Step 6: Set Remotion Composition Duration to Match Audio

// In Root.tsx
import { Composition } from 'remotion';

const fps = 30;
const audioDuration = 18.5; // From ffprobe — use EXACT value

<Composition
  id="ProductDemo"
  component={ProductDemo}
  durationInFrames={Math.ceil(audioDuration * fps)}
  fps={fps}
  width={1920}
  height={1080}
/>

Step 7: Add Audio Track + Sync Scene Transitions

// In ProductDemo.tsx
import { Audio, staticFile, AbsoluteFill, useCurrentFrame, useVideoConfig } from 'remotion';

// Time visual transitions to match voiceover content
const scriptBeats = {
  intro:    { start: 0, end: 3 },      // "Welcome to InstaClaw..."
  problem:  { start: 3, end: 8 },      // "Traditional chatbots..."
  solution: { start: 8, end: 15 },     // "InstaClaw agents take action..."
  cta:      { start: 15, end: 18.5 }   // "Try it today..."
};

export const ProductDemo = () => {
  const frame = useCurrentFrame();
  const { fps } = useVideoConfig();
  const currentTime = frame / fps;

  return (
    <AbsoluteFill>
      {currentTime < scriptBeats.intro.end && <IntroScene />}
      {currentTime >= scriptBeats.problem.start && currentTime < scriptBeats.problem.end && <ProblemScene />}
      {currentTime >= scriptBeats.solution.start && currentTime < scriptBeats.solution.end && <SolutionScene />}
      {currentTime >= scriptBeats.cta.start && <CTAScene />}

      {/* Voiceover track */}
      <Audio src={staticFile('voiceover.mp3')} volume={1.0} />

      {/* Optional: Background music at low volume */}
      <Audio src={staticFile('background-music.mp3')} volume={0.15} />
    </AbsoluteFill>
  );
};

Step 8: Render

npx remotion render ProductDemo output.mp4

Result: Professional MP4 with synchronized voiceover + optional background music.

Workflow 2: Podcast Intro/Outro Generation

# 1. Write intro script
SCRIPT="Welcome to The AI Agent Show, the podcast where we explore how artificial intelligence is reshaping work, creativity, and the future. I'm your host. Let's dive in."

# 2. Generate with professional voice (ElevenLabs for quality)
~/scripts/tts-elevenlabs.sh "$SCRIPT" /tmp/intro-voice.mp3 onyx

# 3. Mix with music bed (music at 20% volume)
~/scripts/audio-toolkit.sh mix /tmp/intro-voice.mp3 public/intro-music.mp3 output/podcast-intro.mp3 0.2

# 4. Normalize final mix
~/scripts/audio-toolkit.sh normalize output/podcast-intro.mp3 output/podcast-intro.mp3

Workflow 3: Audio Summary of Document

When user says: "Summarize this report as audio" or "Read me the highlights"

# 1. Agent summarizes document to ~750 words (~5 min at 150 wpm)
# (Agent writes summary to /tmp/summary.txt)

# 2. Generate audio (OpenAI — cheaper for summaries)
~/scripts/tts-openai.sh "$(cat /tmp/summary.txt)" /tmp/summary-raw.mp3 echo tts-1-hd

# 3. Normalize and compress for mobile
~/scripts/audio-toolkit.sh normalize /tmp/summary-raw.mp3 /tmp/summary-norm.mp3
~/scripts/audio-toolkit.sh compress /tmp/summary-norm.mp3 output/summary.mp3

# 4. Deliver via Telegram (or send as file)
# Agent sends output/summary.mp3 via messaging

Workflow 4: Voice Messages via Telegram

Agent sends voice replies instead of text:

# 1. Agent composes response text
RESPONSE="Here's what I found about your competitor pricing..."

# 2. Generate voice (OpenAI — faster, cheaper for quick messages)
~/scripts/tts-openai.sh "$RESPONSE" /tmp/voice-raw.mp3 nova tts-1

# 3. Convert to Telegram voice format (OGG Opus)
~/scripts/audio-toolkit.sh convert /tmp/voice-raw.mp3 output/voice-reply.ogg opus

# 4. Send as Telegram voice message
# Agent uses messaging tool to send output/voice-reply.ogg

Audio Processing Toolkit

All commands available via ~/scripts/audio-toolkit.sh:

Command	Usage	What It Does
duration	audio-toolkit.sh duration file.mp3	Get exact duration in seconds
normalize	audio-toolkit.sh normalize in.mp3 out.mp3	Normalize volume (loudnorm)
mix	audio-toolkit.sh mix voice.mp3 music.mp3 out.mp3 0.2	Mix voice + music at given volume
compress	audio-toolkit.sh compress in.mp3 out.mp3	Compress for messaging (96k mono)
convert	audio-toolkit.sh convert in.mp3 out.ogg opus	Convert format (mp3/ogg/aac/wav)
trim	audio-toolkit.sh trim in.mp3 out.mp3 2 10	Trim: start at 2s, duration 10s
silence-remove	audio-toolkit.sh silence-remove in.mp3 out.mp3	Remove leading/trailing silence
concat	audio-toolkit.sh concat out.mp3 part1.mp3 part2.mp3	Join multiple files

Voice Selection Guide

Use Case	ElevenLabs Voice	OpenAI Voice	Why
Product demo video	Professional Male/Female	alloy	Authority, trust
Explainer video	Warm Narrator	nova	Approachable, clear
Podcast intro	Deep Professional	onyx	Gravitas
Voice message reply	Casual Conversational	shimmer	Friendly, natural
Document summary	Clear, Measured	echo	Easy to follow
Accessibility reading	Natural, Unhurried	fable	Comfortable pace

OpenAI voice IDs: alloy, echo, fable, onyx, nova, shimmer
ElevenLabs: Use voice IDs from curl -s -H "xi-api-key: $ELEVENLABS_API_KEY" https://api.elevenlabs.io/v1/voices | python3 -m json.tool

Long Text Handling

TTS APIs have per-request character limits. For long content:

1. Check length first: If text > 5000 chars, split into segments
2. Split at sentence boundaries: Never split mid-sentence
3. Generate each segment separately
4. Concatenate with FFmpeg:

# Split text into segments (agent does this in code)
# Generate each: tts-openai.sh "$SEGMENT1" /tmp/part1.mp3
#                tts-openai.sh "$SEGMENT2" /tmp/part2.mp3

# Join them
~/scripts/audio-toolkit.sh concat output/full.mp3 /tmp/part1.mp3 /tmp/part2.mp3 /tmp/part3.mp3

Usage Tracking

ALWAYS track usage. TTS costs add up. Before generating:

# Check if we have budget for this text
python3 ~/scripts/audio-usage-tracker.py check $(echo -n "$TEXT" | wc -c)
# Output: OK — 150000/1800000 chars used (8.3%), 1650000 remaining
# Or:     WARN — 1750000/1800000 chars used (97.2%), consider OpenAI fallback
# Or:     OVER — monthly limit exceeded, switching to OpenAI fallback

# After generating, log it
python3 ~/scripts/audio-usage-tracker.py track $(echo -n "$TEXT" | wc -c) elevenlabs

# Check current status
python3 ~/scripts/audio-usage-tracker.py status

Common Mistakes

1. Not matching video duration to audio. Generate audio FIRST, get exact duration from ffprobe, THEN set Remotion composition duration. Never estimate.

2. Using ElevenLabs for everything. ElevenLabs is premium. Use OpenAI TTS for internal/draft content, voice messages, and document summaries. Reserve ElevenLabs for public-facing content.

3. Ignoring audio normalization. Different TTS providers output at different volumes. Always run audio-toolkit.sh normalize before mixing or delivering.

4. Wrong format for platform. Telegram voice = OGG Opus. Apple = AAC. Web = MP3. Always convert to the right format.

5. Script too long for single API call. ElevenLabs limit is ~5000 chars per call on lower tiers. Split long scripts into segments and concatenate.

6. Not removing TTS silence. Most TTS engines add 0.5-1s silence at start/end. Use silence-remove before syncing to video.

7. Forgetting to track usage. Always run audio-usage-tracker.py track after every generation. A single agent generating hours of audio blows through the budget.

Quality Checklist

Before delivering any audio:

• Audio plays correctly (not corrupted, correct format)
• Volume normalized (loudnorm applied)
• Leading/trailing silence removed
• Correct voice selected for use case
• If Remotion: video duration matches audio duration exactly
• If Remotion: scene transitions align with script beats
• If mixing: background music at 15-25% volume (not overpowering)
• File size appropriate for delivery channel (compressed for Telegram)
• Usage logged via audio-usage-tracker.py track
• Within monthly budget (checked via audio-usage-tracker.py check)
• Output format matches delivery channel (OGG for Telegram, MP3 general, AAC Apple)