Community Discussion · Policy

WeChat hold-to-speak transcription test: Recognition accuracy still lacking

LuguoAug 212026/08/21 284 views

I compared WeChat's voice-to-text feature with the new hold-to-transcribe function and actually tested it out. Conclusion: It depends on the situation. It's suitable for sending long messages when you don't want to disturb others, but not ideal for scenarios where high text accuracy is required.

Policy

3 replies

Ctrl + Enter to reply

Lu JiayiAug 22

Recognition accuracy does tend to crash with long sentences and breathing pauses, but the ability to output text in real-time is useful for asynchronous communication. I work remotely and often need to quickly draft a paragraph to send to group chats. Converting speech to text first and then editing is faster than typing character by character. However, for formal occasions, you still need to manually review it; you can't fully trust it.

Jiang ZhiyuanAug 22

Basic issues like prompt truncation during the grayscale phase suggest that end-to-end monitoring and alerting aren't set up properly. Fluctuations in recognition rates are normal, but have you provided feedback to WeChat regarding recognition results in specific scenarios? I'm curious how their internal release process safeguards this metric.

Pan XuetingAug 21

Haha, I had the same experience last week with that Yinghe Yimai model. It runs smoothly with standard data, but goes haywire when switched to older equipment... Seems like it's not just medical devices; even a crappy voice assistant can't escape this pattern. But let's keep quiet about it—press-and-hold for real-time text output is genuinely convenient. Us impatient folks can't wait those few seconds for transcription...