Narration for a video, a spoken prompt in an app, a subtitle file turned into audio: all of it starts with getting text read out and saved to a file. Windows can do it with what's already installed, and the neural voices are a pip install away. Here's what each route costs you.
Narrator and Edge's Read Aloud use the built-in speech engine, but neither will save what it reads. System.Speech will:
Add-Type -AssemblyName System.Speech
$s = New-Object System.Speech.Synthesis.SpeechSynthesizer
$s.GetInstalledVoices() | ForEach-Object { $_.VoiceInfo.Name + " (" + $_.VoiceInfo.Culture + ")" }
$s.SelectVoice("Microsoft Zira Desktop")
$s.Rate = 0 # -10 .. 10
$s.SetOutputToWaveFile("C:\temp\narration.wav")
$s.Speak((Get-Content "C:\temp\script.txt" -Raw -Encoding UTF8))
$s.Dispose()
No install, no network, and it writes a WAV. The delivery is unmistakably synthetic, so it suits prompts and beeps more than narration.
Add a voice under Settings > Time & language > Speech and it often doesn't appear in GetInstalledVoices(). The two generations of voices live in different registry hives:
Get-ChildItem 'HKLM:\SOFTWARE\Microsoft\Speech\Voices\Tokens'
Get-ChildItem 'HKLM:\SOFTWARE\Microsoft\Speech_OneCore\Voices\Tokens'
On the Windows 11 (25H2) machine I checked, the SAPI hive had two voices and the OneCore hive had a third one that System.Speech could never select. If a voice you installed isn't in the list, this is why.
The voices Edge reads pages with are neural, and edge-tts drives them directly:
pip install edge-tts
edge-tts --text "Hello, world!" --write-media hello.mp3 --write-subtitles hello.srt
edge-tts --list-voices
edge-tts --voice en-US-AriaNeural --file script.txt --write-media narration.mp3
edge-tts --rate=-50% --text "Hello, world!" --write-media slower.mp3
edge-tts --volume=-50% --text "Hello, world!" --write-media quieter.mp3
edge-tts --pitch=-50Hz --text "Hello, world!" --write-media lower.mp3
Options are from the project's README and CLI help. No API key, no account, and it writes MP3 rather than WAV, which matters once a batch of scripts piles up.
A subtitle file states exactly how long each line is on screen. Read that line aloud and it is nearly always longer, because people skim subtitles while a synthesizer pronounces every syllable. Translate first and the gap widens.
cue 1 00:00:00,000 --> 00:00:03,500 (3.5s) speech 3.2s fits
cue 2 00:00:03,500 --> 00:00:07,000 (3.5s) speech 4.6s overruns into cue 3
cue 3 00:00:07,000 --> 00:00:12,000 (5.0s) speech 4.4s fits
What actually works:
Making something new rather than dubbing? Generate the narration first and cut the visuals to its length. Then nothing has to be squeezed.
edge-tts, but they are Microsoft's service, so read the terms before shipping commercial work.
Figures are from September 2026 and these terms change often. "Free to generate, but not for monetised content" is a common combination, and it's the sort of thing that surfaces after the video is published.
The full version, with diagrams and a GUI option for people who don't want a terminal:
Which route do you use for narration?