If you’ve ever tried to localize a video, you know the pain: hiring voice actors, coordinating recording sessions, and then painstakingly syncing the new audio back to the footage. With modern text‑to‑speech (TTS) and voice‑cloning services, you can automate most of that workflow. In this article we’ll build a lightweight AI dubbing pipeline that takes a source video, generates a new voice track in any language you like, and stitches it back together—all with a few lines of Python and a call to ElevenLabs.
TL;DR – Extract audio → transcribe (optional) → generate new speech → align & mux → 🎉
| Item | Reason |
|---|---|
| Python 3.9+ | Main scripting language |
ffmpeg |
Extract and re‑encode audio/video |
pydub |
Simple audio manipulation |
requests |
Call ElevenLabs REST API |
(Optional) whisper or any speech‑to‑text model |
If you want to translate the script before dubbing |
| An ElevenLabs API key | The magic behind high‑quality voice synthesis |
Install the Python dependencies:
pip install pydub requests tqdm
pip install openai-whisper
Make sure ffmpeg is in your PATH – most OS package managers have it (brew install ffmpeg, apt-get install ffmpeg, etc.).
First we need the raw audio track from the video. ffmpeg can do this in a single command:
ffmpeg -i input_video.mp4 -vn -acodec pcm_s16le -ar 44100 -ac 2 original.wav
-vn disables video output.pcm_s16le gives us a lossless WAV file that works well with most TTS services.
If you prefer to keep everything inside Python, subprocess works just as well:
import subprocess
def extract_audio(video_path, wav_path):
cmd = [
"ffmpeg",
"-i", video_path,
"-vn",
"-acodec", "pcm_s16le",
"-ar", "44100",
"-ac", "2",
wav_path,
"-y" # overwrite if exists
]
subprocess.run(cmd, check=True)
extract_audio("input_video.mp4", "original.wav")
If you already have a script, skip this step. Otherwise, you can run an automatic transcription and then translate it with any LLM or translation API. Here’s a quick Whisper example:
import whisper
model = whisper.load_model("base")
result = model.transcribe("original.wav")
script = result["text"]
print("Transcribed script:", script)
You could pipe script into Google Translate, DeepL, or even an OpenAI LLM to get the target language text.
ElevenLabs offers a REST endpoint that can produce lifelike speech in a matter of seconds. The following function sends the text and streams back a WAV file.
import requests
from pathlib import Path
ELEVEN_API_KEY = "YOUR_ELEVENLABS_API_KEY"
VOICE_ID = "EXAVITQu4vr4xnSDxMaL" # default “Rachel” voice – replace with your cloned voice ID
def synthesize(text: str, out_path: Path):
url = f"https://api.elevenlabs.io/v1/text-to-speech/{VOICE_ID}"
headers = {
"xi-api-key": ELEVEN_API_KEY,
"Content-Type": "application/json",
"Accept": "audio/wav"
}
payload = {
"text": text,
"model_id": "eleven_monolingual_v1",
"voice_settings": {
"stability": 0.75,
"similarity_boost": 0.85
}
}
response = requests.post(url, json=payload, headers=headers, stream=True)
response.raise_for_status()
with open(out_path, "wb") as f:
for chunk in response.iter_content(chunk_size=8192):
f.write(chunk)
synthesize(script, Path("dub.wav"))
Pro tip: If you’ve trained a custom voice on ElevenLabs (they call it voice cloning), just replace VOICE_ID with the ID you received after the cloning process. The same API call works for both stock and cloned voices.
The naive approach is to assume the new script has the same duration as the original. In practice, you’ll need to time‑stretch or pad the audio to match the video’s timeline. pydub makes this straightforward:
from pydub import AudioSegment
from pydub.utils import mediainfo
def match_duration(source_wav, target_wav):
src = AudioSegment.from_wav(source_wav)
tgt = AudioSegment.from_wav(target_wav)
src_len = len(src) # milliseconds
tgt_len = len(tgt)
if tgt_len > src_len:
tgt = tgt[:src_len]
else:
silence = AudioSegment.silent(duration=src_len - tgt_len)
tgt = tgt + silence
tgt.export("dub_aligned.wav", format="wav")
return "dub_aligned.wav"
aligned_path = match_duration("original.wav", "dub.wav")
If you need finer control (e.g., per‑sentence alignment), you can use forced‑alignment tools like Aeneas or Montreal Forced Aligner, but that’s beyond the scope of this quick demo.
Now that the audio length matches the video, we simply replace the original track:
ffmpeg -i input_video.mp4 -i dub_aligned.wav -c:v copy -map 0:v:0 -map 1:a:0 -shortest output_dubbed.mp4
-c:v copy tells ffmpeg to keep the video stream unchanged.-shortest ensures the output stops when the shortest stream ends (useful if you have tiny mismatches).
You can wrap this in Python as well:
def mux(video_path, audio_path, out_path):
cmd = [
"ffmpeg",
"-i", video_path,
"-i", audio_path,
"-c:v", "copy",
"-map", "0:v:0",
"-map", "1:a:0",
"-shortest",
out_path,
"-y"
]
subprocess.run(cmd, check=True)
mux("input_video.mp4", aligned_path, "output_dubbed.mp4")
That’s it—your video now speaks in a brand‑new voice!
Putting everything together in a single script makes it reusable:
def dub_video(
video_file: str,
target_text: str,
output_file: str,
voice_id: str = VOICE_ID,
):
extract_audio(video_file, "original.wav")
synthesize(target_text, Path("dub.wav"))
aligned = match_duration("original.wav", "dub.wav")
mux(video_file, aligned, output_file)
dub_video(
video_file="input_video.mp4",
target_text=script, # could be the original script or a translated version
output_file="final_dub.mp4"
)
Feel free to expand it with command‑line arguments, batch processing, or a tiny Flask API if you want a web‑based dubbing service.
<phoneme alphabet="ipa" ph="..."> to force correct pronunciation.
You’ve just built a functional AI dubbing tool using open‑source utilities and the powerful ElevenLabs TTS platform. From extracting audio to syncing the final track, the whole workflow can be scripted in under a minute per video, opening doors to rapid localization, prototype voice‑overs, or even personalized video messages.
Ready to give your videos a fresh voice? Grab your ElevenLabs API key, follow the steps above, and start dubbing today! 🚀
Try ElevenLabs now and see how easy high‑quality voice cloning can be: https://try.elevenlabs.io/kr07zfuqn1bp.