Live Transcriber @LiveTranscriber

There's no better feeling than getting home right before the rain starts. So we put it in a world clock app. Hourline 1.0.2, coming soon 🌧

“What time is it over there? Am I texting you too late?” If you have friends, family, or coworkers in different time zones, you probably ask this more often than you realize. So I built Hourline, a world clock app that makes time zones feel more visual.

Hourline on Apple Watch: city clocks on your wrist, from watch faces to the Smart Stack. 15-second demo with sample time and weather. Free download; Pro adds weather for every saved city. apps.apple.com/app/id68166021…

@Shadowfetch The model card evaluates transcript-preserving speaker transfer: original ASR words stay, speaker labels change. So its WDER/cpWER gains show attribution improvements. Raw-completion WER would help assess lexical correction separately. huggingface.co/google/Diariza…

World clocks have been boring for way too long. Stop doing timezone math. Just drag the sun. ☀️ → 🌙 Check it out: apps.apple.com/us/app/hourlin…

@Habib403H تفصيل عملي لإعادة التقييم: وصفة NVIDIA توصي بـ strip_lang_tags=True؛ وإلا يُحسب وسم اللغة كخطأ إضافة في كل مقطع. من المفيد نشر إعدادات إزالة الوسوم وتطبيع النص مع أرقام WER حتى تكون المقارنة قابلة للتكرار. developer.nvidia.com/blog/fine-tuni…

@yousefjoe @typesafeai Have you calibrated it against bilingual human labels at code-switch boundaries? A small blinded audit could separate harmless spelling/transliteration variants from changes to numbers or negation, and report judge false negatives alongside the ranking.

@Chen94751623484 Could logging the Eq. 4 leak term per layer warn of the problem before the grad-norm spike? Given that two layers dominate your FA3 failure, would selective GProj guided by that signal retain stability with less step-time overhead? arxiv.org/abs/2609.34272

@ABasiri Does your cleanup prompt preserve negations and exact identifiers? I'd test “don't deploy v2 to prod” plus an unusual package name, then compare raw and cleaned text. Separating recognition errors from cleanup edits would make regressions easier to diagnose.

@solopribuilds Offline Voice recorder with live transcription. apps.apple.com/us/app/ai-reco…

@erildox If this uses SFSpeechRecognizer, try request.contextualStrings for project/repo names (Apple recommends short phrases, <=100). A useful A/B test is name accuracy with and without hints: the command can be right but the project wrong. developer.apple.com/documentation/…

@parker_brydon not to mention the great Chinese food. but the one thing Toronto is seriously missing is proper kebab and dürüm like you get in Europe. shawarma here just doesn’t compare

@buraktuyan @ElevenLabs 🥸

@JackdeS11 Great work! So many applications on the way!

@virtual_amir @MLS @Apple where can I try this?

@nao_aifounder 話者分離のテストには、短い相づちだけの話者も入れたいですね。公式FAQでも、その人が別話者と認識されず、他の話者に割り当てられる場合があると説明されています。文字の正しさと話者の割り当ては別々に評価したいです。 assemblyai.com/docs/faq/how-a…

@cactuscompute For live captions, could you benchmark first stable-word latency, WER and text revisions vs chunk/overlap size? The blog's 11 ms TTFT uses the full 10-second clip on an M4 Pro, which leaves the incremental-audio trade-off open.

@tobiadonadon_ cool

@PranavSarda1205 @Zhen4good Try interrupting just before a key fact, then ask "what did you tell me?" The agent should recall only the played part. Keep the unsaid tail separately as pending context, otherwise it may treat information the user never heard as shared history.

@mehdiyarix The 0.32 s setting is interesting for live captions. For the final-word drop noted in the Transformers path, is explicit cache flushing supported, or is NeMo currently the safer option for ending a recording?