The second time you send a voice message the waveform lags reality by ~1 second

### Steps to reproduce

1. Send a VM on a fresh Element instance, as per https://github.com/vector-im/element-web/issues/22584
2. Send another one - e.g. say "1, 2, 3, 4, 5"
3. Rejoice that you get a waveform preview!
4. Hit send
5. Play it back, and discover that the waveform lags reality by about 1 second, and so the last second is truncated - for instance, it will show peaks for only "1, 2, 3, 4"

### Outcome

#### What did you expect?

Waveform should not be truncated and match the reality of the message

#### What happened instead?

Waveform lags the reality by 1s and is truncated., so you can't reliably scrub within it, and it doesn't show an accurate representation of the message contents.

### Operating system

macOS 12

### Browser information

Chrome 102

### URL for webapp

pr8843--matrix-react-sdk.netlify.app

### Application version

pr8843

### Homeserver

matrix.org

### Will you send logs?

No