Speech & Audio

egaki handles text-to-speech, voice cloning, audio stem separation, and BPM detection.

Text-to-speech

1egaki speech "Hello, this is a test." -o hello.mp3

Specify a model and voice:

1234egaki speech "Welcome to the future of video." \
  --model sonic-3.5 \
  --voice <voice-id> \
  -o narration.mp3

Read from stdin

Pipe text from a file:

1cat script.txt | egaki speech --stdin -o narration.mp3

Speed control

Cartesia models support --speed from 0.6 to 1.5:

1egaki speech "Speak faster." --speed 1.3 -o fast.mp3

Available providers

Provider	Models	Notes
OpenAI	`tts-1`, `tts-1-hd`	Standard quality
Cartesia	`sonic-3.5`, `sonic-3`	Best quality, speed control
ElevenLabs	`eleven_v3`, `eleven_multilingual_v2`, `eleven_flash_v2_5`	Multilingual

Voice cloning

Clone a voice from an audio clip and get a reusable voice ID:

1egaki voice clone recording.mp3 --name "my-voice" --json

Workflow: clone from a song

Separate vocals first, then clone:

12345678# 1. Separate vocals
egaki demucs song.mp3 --stems vocals -o stems/

# 2. Clone the isolated voice
egaki voice clone stems/song-vocals.mp3 --name "singer" --json

# 3. Generate speech with the cloned voice
egaki speech "Your text here." --voice <voice-id> -m sonic-3.5 -o output.mp3

Providers

Cartesia (default): instant cloning, up to 10s of audio, free
ElevenLabs: longer clips, --remove-background-noise option

12345egaki voice clone noisy-audio.mp3 \
  --provider elevenlabs \
  --name "clean-voice" \
  --remove-background-noise \
  --json

Audio stem separation (demucs)

Separate a song into individual stems using fal.ai's Demucs model:

1egaki demucs song.mp3 --stems vocals,other -o stems/

Available stems

vocals, drums, bass, other, guitar, piano

Models

Model	Stems	Best for
`htdemucs`	vocals, drums, bass, other	General purpose
`htdemucs_ft`	vocals, drums, bass, other	Fine-tuned, higher quality
`htdemucs_6s` (default)	vocals, drums, bass, other, guitar, piano	6-stem separation

1egaki demucs song.mp3 --model htdemucs_6s --stems vocals,guitar -o stems/

BPM detection

Detect the tempo of an audio file locally. No API call needed; runs in milliseconds using peak detection and interval analysis.

1egaki bpm song.mp3

Output:

12BPM: 120
Interval: 0.500s

Use --json for structured output:

12egaki bpm track.wav --json
# { "bpm": 120, "intervalInSeconds": 0.5 }

Pipe audio from another command:

1ffmpeg -i video.mp4 -f mp3 - | egaki bpm --stdin

Beat-synced videos

Use the detected BPM in your MDX frontmatter to align scene cuts to the music:

12345# 1. Detect the BPM
egaki bpm soundtrack.mp3 --json
# { "bpm": 129 }

# 2. Use it in your video

12345678910---
fps: 30
bpm: 129
---

<Audio src="/soundtrack.mp3" />

# Verse duration=8beats

Scene cuts land on the beat.

Optional dependency

BPM and loudness detection both require node-web-audio-api, a native audio processing library. It's listed as an optional dependency; if missing, the command prints install instructions:

1pnpm add node-web-audio-api

Loudness measurement (LUFS)

Measure perceived loudness in LUFS using K-weighting and two-pass gating inspired by EBU R128. Runs locally, no API call needed.

1egaki loudness song.mp3

Output:

1234Integrated: -14.2 LUFS
Max: -8.1 LUFS
Min: -28.4 LUFS
Range: 20.3 LU

Use --json for structured output:

1egaki loudness narration.mp3 --json

What the numbers mean

Metric	Description
Integrated	Average perceived loudness across the whole file (gated)
Max	Loudest 400ms block
Min	Quietest 400ms block above the gate threshold
Range	Dynamic range (max - min) in LU

Common loudness targets

Platform	Target LUFS
YouTube	-14
Spotify	-14
Apple Music	-16
Podcast	-16 to -18
Broadcast TV	-24

Compare your audio against these targets to know if it needs volume adjustment before using it in a video.