PROMO REEL WITH LOCAL MAC VOICEOVER
Checked October 9, 2026
Article: /blog/resources/ai-promo-reel-local-voiceover-mac

This original short fixture verifies Murmur CLI speech, MCP speech/job retrieval,
and a local FFmpeg vertical render. Claude Opus 5.5 and GPT-6 Astra are suggested
assistant choices for adapting the prompt; neither was run for this fixture.
Murmur generates voice audio. The assistant writes/plans code and assets.

REQUIREMENTS
Apple Silicon Mac, macOS 15+, licensed Murmur with setup complete and local
automation enabled. Install CLI in Murmur Settings > Automation.
The video helper separately needs ffmpeg, ffprobe, Python 3 and Pillow.
Test environment: development build reporting Murmur 1.0.18; FFmpeg 9.0.2;
Pillow 11.3.0; Apple M3 Pro. Released-binary parity was not verified.
If these video dependencies are absent, configure them separately or use your
existing video editor. They are not prerequisites for Murmur's managed TTS runtime.
Use a fresh project folder, and pick an installed model/voice from your own Mac.

1. CHECK READINESS

murmur --help
murmur status --json
murmur models --json
murmur voices --model qwen3-base --json
ffmpeg -version
ffprobe -version
python3 -c "import PIL; print(PIL.__version__)"

The fixture used qwen3-base / Ryan. The Mac also exposed the English Aiden
preset. Kokoro was unavailable on this Mac, so do not assume af_bella works
with the current default model. Never change defaults just to run this fixture.

2. CREATE ORIGINAL SCRIPT IN A NEW PROJECT FOLDER

mkdir promo-reel-01
cd promo-reel-01
cat > script.txt <<'SCRIPT'
Start with one clear idea. Write a short script. Generate the voice on your Mac, then build the motion around it.
SCRIPT

3. GENERATE A LOCAL WAV

murmur generate --input script.txt --model qwen3-base --voice Ryan \
  --language EN-US --speed 1.0 --output narration.wav --json

The original result was succeeded, with Qwen3-TTS Base / Ryan, 7.12 seconds,
mono 24,000 Hz PCM signed 16-bit WAV, 341,804 bytes. Measure your own result.
No --force is used. For a revision choose a new output filename.

Optional detached request (separate new take; not required for the render):

murmur generate --input script.txt --model qwen3-base --voice Ryan \
  --output narration-v2.wav --detach --json
murmur jobs get JOB_ID --json

Replace JOB_ID with the returned UUID. Wait for succeeded plus an artifact path.

4. SAVE THE ORIGINAL TITLE-CARD / FFMPEG HELPER

cat > make-reel.py <<'PY'
from PIL import Image, ImageDraw, ImageFont
from pathlib import Path
import subprocess, json, math

root = Path.cwd()
targets = [root / f'slide-{n}.png' for n in range(1, 4)]
targets += [root / name for name in ('reel-smoke-test.mp4', 'render-command.json', 'output-metadata.json')]
if any(path.exists() for path in targets):
    raise SystemExit('Output exists. Use a fresh folder or new filenames for this version.')
font_path = '/System/Library/Fonts/Supplemental/Arial.ttf'
font = ImageFont.truetype(font_path, 86)
small = ImageFont.truetype(font_path, 42)
slides = [
    ('WRITE THE SCRIPT', 'Start with one clear idea.'),
    ('GENERATE THE VOICE', 'Local narration on your Mac.'),
    ('BUILD THE MOTION', 'Let the audio set the pace.')
]
for n, (title, subtitle) in enumerate(slides, 1):
    im = Image.new('RGB', (1080, 1920), (17, 24, 39))
    d = ImageDraw.Draw(im)
    d.rounded_rectangle((84, 500, 996, 1180), radius=44, fill=(24, 33, 54))
    d.rectangle((84, 500, 996, 516), fill=(145, 162, 255))
    d.text((118, 584), f'0{n}', font=font, fill=(145, 162, 255))
    for j, word in enumerate(title.split()):
        d.text((118, 742 + j*110), word, font=font, fill=(245, 247, 255))
    d.text((118, 1110), subtitle, font=small, fill=(191, 199, 224))
    d.text((118, 1350), 'Murmur workflow smoke test', font=small, fill=(145, 162, 255))
    im.save(root / f'slide-{n}.png')

audio = root / 'narration.wav'
metadata = json.loads(subprocess.check_output([
    'ffprobe', '-v', 'error', '-show_format', '-show_streams', '-of', 'json', str(audio)
], text=True))
duration = float(metadata['format']['duration'])
parts = math.ceil(duration * 30 / 3) / 30
inputs = []
for n in range(1, 4):
    inputs += ['-loop', '1', '-framerate', '30', '-i', str(root / f'slide-{n}.png')]
filters = []
for n in range(3):
    filters.append(
        f"[{n}:v]zoompan=z='min(zoom+0.00035,1.03)':d=1:s=1080x1920:fps=30,"
        f"trim=duration={parts},fade=t=in:st=0:d=0.15,"
        f"fade=t=out:st={parts-0.15}:d=0.15,setpts=PTS-STARTPTS[v{n}]"
    )
filters.append('[v0][v1][v2]concat=n=3:v=1:a=0,format=yuv420p[v]')
command = [
    'ffmpeg', '-hide_banner', '-loglevel', 'warning', '-n', *inputs,
    '-i', str(audio), '-filter_complex', ';'.join(filters),
    '-map', '[v]', '-map', '3:a:0', '-t', str(duration),
    '-c:v', 'libx264', '-preset', 'fast', '-crf', '23',
    '-c:a', 'aac', '-b:a', '192k', '-ar', '48000',
    '-movflags', '+faststart', str(root / 'reel-smoke-test.mp4')
]
(root / 'render-command.json').write_text(json.dumps(command, indent=2))
subprocess.run(command, check=True)
output = json.loads(subprocess.check_output([
    'ffprobe', '-v', 'error', '-count_frames', '-show_format', '-show_streams',
    '-of', 'json', str(root / 'reel-smoke-test.mp4')
], text=True))
(root / 'output-metadata.json').write_text(json.dumps(output, indent=2))
print(json.dumps(output, indent=2))
PY

python3 make-reel.py

This uses the same FFmpeg arguments/filters as the measured fixture. The helper
uses current-folder paths instead of the test's temporary folder. The helper checks every generated filename before creating cards, and FFmpeg -n
refuses to replace a movie. For another version use a new folder or change all
output filenames. Existing audio is used as input and is never replaced.

5. CHECK THE FINAL MP4

ffprobe -v error -count_frames -show_format -show_streams \
  -of json reel-smoke-test.mp4
ffmpeg -hide_banner -loglevel error -i reel-smoke-test.mp4 -f null -

Original measured video: 1080x1920, 30 fps, H.264 High/yuv420p, AAC mono
48,000 Hz, 214 video frames, 7.133333 seconds, 210,042 bytes. Full video/audio
decode returned 0. A midpoint frame was visually inspected. The title cards are
summaries, not timed speech captions. Listen to the completed movie, confirm the
words, review timing, and add checked captions before using it as a public reel.
This test does not independently transcribe every spoken word or measure audience
response. Your duration and size may differ; use the actual metadata.

6. MCP OPTION: LET AN AGENT REQUEST AUDIO

For a normal app installation, the documented Claude Code registration is:

claude mcp add --transport stdio --scope user murmur -- \
  /Applications/Murmur.app/Contents/Helpers/murmur mcp serve

For Codex:

codex mcp add murmur -- /Applications/Murmur.app/Contents/Helpers/murmur mcp serve

Open the agent in this project folder. Call murmur_status, murmur_list_models,
and murmur_list_voices before generating. The MCP server limits file access to
the directory where it launches. Typical tool-call parameters:

{
  "name": "murmur_generate_speech",
  "arguments": {
    "text": "Build a reel with local narration.",
    "output_path": "mcp-take.wav",
    "model_id": "qwen3-base",
    "voice_id": "Ryan",
    "language": "EN-US",
    "speed": 1.0
  }
}

This exact local tool invocation was tested and returned a queued job. Poll with:

{
  "name": "murmur_get_job",
  "arguments": { "job_id": "RETURNED_JOB_UUID" }
}

The original completed artifact was Qwen3-TTS Base / Ryan, 2.24 seconds.
MCP generation is asynchronous by default; "wait": true requests completion
before returning. A queued response alone is not audio completion.

Batch shape verified in the installed tool schema, not executed for this fixture:

{
  "name": "murmur_generate_batch",
  "arguments": {
    "items": [
      { "id": "hook", "text": "Start with one clear idea.",
        "output_path": "audio/01-hook.wav", "model_id": "qwen3-base",
        "voice_id": "Ryan", "speed": 1.0 },
      { "id": "result", "text": "Build the motion around your voiceover.",
        "output_path": "audio/02-result.wav", "model_id": "qwen3-base",
        "voice_id": "Ryan", "speed": 1.0 }
    ]
  }
}

CLI batch manifests instead use camelCase field names, such as outputPath,
modelID and voiceID, then: murmur batch --manifest scenes.json --json

7. EDITABLE AGENT PROMPT

Create a vertical promo reel for [product] aimed at [viewer]. Show [one verified
result] and end with [one next step]. First write a short script and a three-scene
plan. Keep narration separate from on-screen text. Use only the screenshots and
original assets in this workspace. Check Murmur status, installed models, and
preset voices before generating. Use one model and voice across scenes, save new
WAV takes under audio/, measure the finished audio, and time the visuals from
those measurements. Render a 1080x1920 MP4 at 30 fps. Return the script, editable
source, audio paths, render command, and verified output metadata. Keep
publication for my review.

Adapt that prompt to your agent host and selected coding model. We checked the
provider pages for Claude Opus 5.5 and GPT-6 Astra naming, but did not run either
for this sample. The host needs local file/terminal access or Murmur MCP tools.
Local TTS has no per-character fees on a licensed Murmur installation. Your
assistant plan/API charges and finite Mac resources remain separate.

CHECKED PRIMARY REFERENCES
https://www.murmurtts.com/docs/automation
https://ffmpeg.org/ffmpeg.html
https://ffmpeg.org/ffmpeg-filters.html
https://www.anthropic.com/claude-opus-5-5
https://developers.openai.com/api/docs/models/gpt-6-astra