Search
71 results for “bgta”
-
A #Leadtech estem buscant un/a 𝘿𝙚𝙫𝙊𝙥𝙨 𝙀𝙣𝙜𝙞𝙣𝙚𝙚𝙧 per unir-se a l'equip de Manuel Trujillo, el nostre 𝙎 𝙈𝙖𝙣𝙖𝙜𝙚𝙧.
Fes-li una ullada als requisits indispensables de la mà de Raúl ( @bgta )
https://twitter.com/leadtech_social/status/1714607212620841397?t=9xa9Jkd0gcBvs32OPcg0XA&s=19
-
I Wanted Podcast Transcriptions. iOS 26 Delivered (and Nearly Melted My Phone).
Testing iOS 26’s on-device speech recognition: faster than realtime, but your phone might disagree
Apple’s iOS 26 introduced
SpeechTranscriber– a promise of on-device, private, offline podcast transcription. No cloud, no subscription, just pure silicon magic. I built it into my RSS reader app. Here’s what actually happened.The Setup
- Device: iPhone 17 Pro Max (Orange, if you’re curious)
- iOS Version: 26.2
- Test Episodes:
- The Talk Show #436 (95 minutes)
- Upgrade #594 (106 minutes)
- ATP #668 (114 minutes)
- ATP #669 (122 minutes)
The Good News: It’s Actually Fast
EpisodeDurationTranscription TimeRealtime FactorWordsWords/secTalk Show #4361h 35m15m 22s6.2x17,30318.8Upgrade #5941h 46m20m 4s5.3x19,97516.6ATP #6681h 54m24m 49s4.6x23,89216.04.6x to 6.2x faster than realtime. Nearly 2-hour podcasts transcribed in under 25 minutes. The Neural Engine absolutely crushes this.
The Pipeline Breakdown
The transcription happens in two phases (example from Upgrade #594):
- Audio Analysis: 2m 2s
- Initial pass through the audio file
- Roughly 1 second of analysis per minute of audio
- Results Collection: 18m 0s
- Iterating through ~1,288 speech segments
- Each segment yields transcribed text
The Bad News: Thermal Throttling Is Real
During my first test, I made a critical mistake: running two transcriptions simultaneously while charging.
The result? My phone got noticeably hot. Battery optimization warnings appeared. And performance dropped dramatically:
ConditionRealtime FactorPerformance HitSingle transcription4.6x – 6.2xBaselineTwo parallel transcriptions2.7x46% slowerThe logs showed alternating progress updates as iOS juggled both workloads:
🎙️ 📝 Progress: 34% - 88 segments // Transcription A 🎙️ 📝 Progress: 44% - 98 segments // Transcription B 🎙️ 📝 Progress: 37% - 98 segments // Transcription AThe Neural Engine throttles hard when thermals get bad. When I ran a single transcription without charging, the ETA stayed consistent and completed on schedule.
The Ugly: iOS Kills Background Tasks
Even with
BGTaskScheduler, iOS terminated my background transcription:🎙️ Background transcription task triggered by iOS ⏱️ Background transcription task expired (iOS terminated it)For long podcasts, you need to keep the app in foreground. iOS’s aggressive app suspension doesn’t play nice with hour-long ML workloads.
AI Chapter Generation: The Real Win
Here’s where it gets interesting. Once you have a transcript, generating AI chapters is blazingly fast.
Note: ATP, Talk Show, and Upgrade already include chapters via ID3 tags – this is an experiment to see what on-device AI can generate. But Planet Money doesn’t have chapters, making it a real use case where AI generation adds genuine value.
And we’re not alone in this approach. As Mike Hurley and Jason Snell discussed on Upgrade #594, Apple is doing exactly this in iOS 26.2’s Podcasts app:
“One of the most interesting things to me is the changes in the podcast app in 26.2… AI generated chapters for podcasts that do not support them… They are creating their own chapters based on the topics.”
Jason nailed the insight: “The transcripts [are] a feature that unlocks a lot of other features, because now they kind of understand the content of the podcast.”
That’s exactly what we’re doing here – using on-device transcription as a foundation for AI-powered chapter generation:
EpisodeTranscript SizeChapters GeneratedTimeATP #669143,603 chars (~26,387 words)27 chapters2m 1sTalk Show #436~17,303 words13 chapters1m 40sThe AI identified topic changes, extracted key phrases for timestamps, and generated descriptive chapter titles – all in under 2 minutes for multi-hour podcasts.
Sample generated chapters:
📍 0:00-2:18: Snowfall in Richmond 📍 42:43-49:11: Intel-Apple Chip Collaboration Speculations 📍 62:46-65:00: Executive Transitions at Apple 📍 95:56-105:04: Core Values and Apple's EvolutionThe Code
Using iOS 26’s
SpeechTranscriberis surprisingly clean:@available(iOS 26.0, *) func transcribe(fileURL: URL) async throws -> String { let locale = try await findSupportedLocale(preferring: "en") let transcriber = SpeechTranscriber(locale: locale, preset: .transcription) let analyzer = SpeechAnalyzer(modules: [transcriber]) let audioFile = try AVAudioFile(forReading: fileURL) if let lastSample = try await analyzer.analyzeSequence(from: audioFile) { try await analyzer.finalizeAndFinish(through: lastSample) } var transcription = "" for try await result in transcriber.results { if result.isFinal { transcription += String(result.text.characters) + " " } } return transcription }Fast vs Accurate Mode: A Surprising Finding
iOS 26 offers two main transcription presets:
.transcription– Standard accurate mode.progressiveTranscription– “Fast” mode with progressive results
I assumed Fast mode would be… faster. The results were mixed.
EpisodeModeConditionRealtime FactorWords/secTalk Show #436AccurateSolo, cold6.2x18.8Upgrade #594AccurateSolo5.3x16.6ATP #668AccurateSolo4.6x16.0Planet MoneyFastSolo3.8x12.2Planet MoneyAccurateSolo, warm3.5x11.4On the same 31-minute episode, Fast mode (3.8x) was slightly faster than Accurate (3.5x). But both were significantly slower than the longer episode tests – likely due to residual heat from previous runs.
The “progressive” preset appears optimized for live/streaming transcription. For batch processing of pre-recorded files, results are similar when thermals are equivalent.
Lesson: Don’t assume “fast” means faster for your use case. Profile both.
Recommendations
- Use
.transcriptionfor downloaded files – It’s actually faster for batch processing - Don’t charge while transcribing – Thermal throttling is real
- One transcription at a time – The Neural Engine doesn’t parallelize well
- Keep the app in foreground – iOS will kill background ML tasks
- Expect ~5x realtime – About 12-13 minutes per hour of audio under ideal conditions
The Verdict
iOS 26’s on-device transcription is genuinely impressive:
- Privacy: Audio never leaves your device
- Speed: 5x faster than realtime (when not throttled)
- Quality: Surprisingly accurate for conversational podcasts
- Offline: Once the model is downloaded, no internet required
The main gotchas are thermal management and iOS’s background task limitations. But for a first-generation on-device transcription API? Apple’s Neural Engine delivers.
Now if you’ll excuse me, I have 26,387 words of ATP to search through.
Tested on iPhone 17 Pro Max running iOS 26.x. Your mileage may vary on older devices.
Raw Test Data
Upgrade #594
- Audio Duration: 1h 46m 24s (106 min)
- Audio Analysis Phase: 2m 2s
- Results Collection Phase: 18m 0s
- Total Transcription Time: 20m 4s
- Realtime Factor: 5.3x (faster than audio playback)
- Words Transcribed: 19,975
- Processing Rate: 16.6 words/sec
- Segments Processed: 1,288
ATP #668
- Audio Duration: 1h 53m 54s (114 min)
- Audio Analysis Phase: 2m 20s
- Results Collection Phase: 22m 28s
- Total Transcription Time: 24m 49s
- Realtime Factor: 4.6x (faster than audio playback)
- Words Transcribed: 23,892
- Processing Rate: 16.0 words/sec
- Segments Processed: 1,557
ATP #669 Chapter Generation
- Audio Duration: 2h 2m 13s (122 min)
- Transcription Size: 143,603 characters, ~26,387 words
- Chapters Generated: 27
- Total Time: 2m 1s
- Processing Rate: ~219 words/sec
Talk Show #436
- Audio Duration: 1h 35m 52s (95 min)
- Audio Analysis Phase: 1m 37s
- Results Collection Phase: 13m 44s
- Total Transcription Time: 15m 22s
- Realtime Factor: 6.2x (faster than audio playback) ← Fastest test!
- Words Transcribed: 17,303
- Processing Rate: 18.8 words/sec
- Segments Processed: 971
Talk Show #436 Chapter Generation
- Transcription Size: ~17,303 words
- Chapters Generated: 13
- Total Time: 1m 40s
Planet Money – Chicago Parking Meters (Fast Mode)
- Audio Duration: 30m 56s (31 min)
- Audio Analysis Phase: 1m 3s
- Results Collection Phase: 7m 5s
- Total Transcription Time: 8m 9s
- Realtime Factor: 3.8x
- Words Transcribed: 5,981
- Processing Rate: 12.2 words/sec
- Segments Processed: 472
- Mode:
.progressiveTranscription(Fast)
Planet Money Chapter Generation (Fast Mode)
- Transcription Size: ~5,981 words
- Chapters Generated: 8
- Total Time: 31.9 sec
Planet Money – Accurate Mode (Parallel Stress Test)
- Audio Duration: 30m 56s (31 min)
- Audio Analysis Phase: 1m 9s
- Results Collection Phase: 10m 8s
- Total Transcription Time: 11m 19s
- Realtime Factor: 2.7x ← Severely throttled (ran 2 simultaneous)
- Words Transcribed: 5,983
- Processing Rate: 8.8 words/sec
- Segments Processed: 476
- Mode:
.transcription(Accurate) - Note: Ran in parallel with another transcription – 46% performance hit
Planet Money – Accurate Mode (Solo, Warm Device)
- Audio Duration: 30m 56s (31 min)
- Audio Analysis Phase: 1m 11s
- Results Collection Phase: 7m 32s
- Total Transcription Time: 8m 44s
- Realtime Factor: 3.5x ← Device still warm from previous tests
- Words Transcribed: 5,983
- Processing Rate: 11.4 words/sec
- Segments Processed: 477
- Mode:
.transcription(Accurate) - Note: Slightly slower than Fast mode on same episode (thermal impact)
Device Observations
- Thermal: Significant heat when running multiple transcriptions while charging
- Thermal Carryover: Running tests back-to-back shows degraded performance (6.2x cold → 3.5x warm)
- Cool-down Recommended: Wait 5-10 minutes between long transcriptions for optimal performance
- Battery Notifications: Battery optimization warnings triggered during parallel operations
- Background Tasks: iOS terminated BGTaskScheduler tasks during long transcriptions
- Beta Warning:
Cannot use modules with unallocated locales [en_US (fixed en_US)]– appears in logs but doesn’t block functionality
-
Obre al #Gòtic el #GuitarLegendsHall, un museu amb les guitarres de grups mítics de #rock https://beteve.cat/cultura/obre-museu-guitar-legends-hall-barcelona/
-
La reencarnació elèctrica del mític #R5 dels anys 80 es vendrà a finals d'any
https://www.ccma.cat/324/la-reencarnacio-electrica-del-mitic-r5-dels-anys-80-es-vendra-a-finals-dany/noticia/3278814/ -
Would you like to kick-start your professional career in #Leadtech? This is your chance!
Organic Marketing Director, 𝙑𝙞𝙠𝙮, is looking for a 𝙎𝙀𝙊 𝙄𝙣𝙩𝙚𝙧𝙣.
What’re you waiting for? Click and apply ➡️ cutt.ly/Ywtt8InE
https://twitter.com/leadtech_social/status/1676539499373907970?t=YwWDCk_AAr0lumyBBuo-Lg&s=09
-
Busquem #Sysadmins..ets #Sysadmin? What are you waiting for? #Leadtech
PD: Quin xicot mes aixerit que surt...
https://twitter.com/leadtech_social/status/1668942043161272321?t=O-O0fyoh7CvFLyhnbSIf1w&s=09
-
🚀 Les nostres companyes de #Leadtech Cristina Gil, Reputation Specialist, i Maria Corina Redondo, Social Media Specialist, van decidir utilitzar el seu pressupost anual de formació per assistir al #InternationalSocialSummit a #Barcelona. I tu, en què vols formar-te aquest any? 🤔 #Talent
-
Hem fet un vídeo a la feina parlant de com ens organitzem el nostre horari gràcies a la flexibilitat que tenim a @leadtech_social ¿A què no endevineu qui surt?
Perquè treballar així dona gust ;)
---
RT @leadtech_social
𝙏𝙪 𝙩𝙞𝙚𝙢𝙥𝙤, 𝙩𝙪 𝙧𝙞𝙩𝙢𝙤: 𝙥𝙧𝙤𝙙𝙪𝙘𝙩𝙞𝙫𝙞𝙙𝙖𝙙 𝙖𝙡 𝙢𝙖́𝙭𝙞𝙢𝙤⏱️Sean como sean tus días, en #Leadtech ofrecemos flexibilidad horaria. ¿Quieres saber cómo le sacan provecho nuestros/as leadtechers? ⬇️
#FlexibilidadHoraria #Equipo…
https://twitter.com/leadtech_social/status/1654090547563200512 -
Quin noi més eixerit hi ha al capdavant de la taula... 😂
---
RT @leadtech_social
💻 “𝙄𝙣𝙛𝙤𝙧𝙢𝙖𝙩𝙞𝙤𝙣 𝙨𝙚𝙘𝙪𝙧𝙞𝙩𝙮 𝙞𝙨 𝙖 𝙥𝙧𝙤𝙘𝙚𝙨𝙨, 𝙣𝙤𝙩 𝙖 𝙥𝙧𝙤𝙙𝙪𝙘𝙩.” —𝘽𝙧𝙪𝙘𝙚 𝙎𝙘𝙝𝙣𝙚𝙞𝙚𝙧.
We’re looking for an Information Security Specialist to join our IT team at #Leadtech.
Apply now or find more information here: ▶️ https://cutt.ly/t8hKfwa
https://twitter.com/leadtech_social/status/1633827397261983744 -
#Warframe és a punt de canviar, i si et costava començar-ho, és la teva oportunitat: té nou protagonista, món i història
https://www.3djuegospc.com/mmo/warframe-esta-a-punto-cambiar-te-costaba-empezarlo-tu-oportunidad-tiene-nuevo-protagonista-mundo-e-historia -
Sabeu qui era l'avatar de #Division? 😜
---
RT @leadtech_social
🪅En #Leadtech este carnaval nos hemos trasladado al 𝙢𝙪𝙡𝙩𝙞𝙫𝙚𝙧𝙨𝙤, donde sucede “todo a la vez en todas partes”. Algunos lo celebraron desde la oficina, mientras otros lo hicieron desde diversos puntos de España.¿Adivinas quiénes fueron los ganadores del concurso? ⬇️
https://twitter.com/leadtech_social/status/1628729848783855619 -
Ja el tenim aquí! Feia temps que volia renovar el portàtil i gràcies a una oferta d'#Amazon finalment ha caigut el #Leonovo #Legion 5 #Gen6
Més content que un gínjol 😜
https://www.lenovo.com/es/es/laptops/legion-laptops/legion-5-series/Legion-5-15ACH6H/p/88GMY501582
-
Sembla que #Ubisoft está transferint a la galeria de #UbisoftConnect Tota els jocs que tenies comprats a #Stadia
A mi em surten uns quants i puc jugar-los a #GeForceNow
Ref: https://twitter.com/VaDeCloud/status/1604394443670720512?t=nJgtJKUWiqQ-3UsZJ3Mytw&s=19
-
Bona nit #Tenno … Hora de dormir 😴
-
Mama! Que surto a la tele!
---
RT @leadtech_social
¡𝙁𝙚𝙡𝙞𝙯 #𝘿𝙞́𝙖𝙙𝙚𝙡𝙖𝙄𝙣𝙛𝙤𝙧𝙢𝙖́𝙩𝙞𝙘𝙖 a todas las personas que utilizáis, necesitáis y amáis esta ciencia! 🎉Desde Leadtech queremos agradecer a tod@s l@s que nos acercáis a personas que están a miles de kilómetros de distancia con este vídeo ⬇️
#Leadtech #Team
https://twitter.com/leadtech_social/status/1601169298701307904 -
¿Vols formar part del nostre equip? A #Leadtech estem contractant!
#Hiring #Job #IT #SysAdmin #RemoteWork
https://www.linkedin.com/jobs/view/3362518028/?refId=jA6MnvmWwaoUWlEI5pz2ZQ%3D%3D
-
Acabo de trobar una altra manera de certificar el teu perfil a #Mastodon
-
Apassionat de les noves tecnologies, ment autodidacta i inquieta. Estic al món de la informàtica des dels 12 anys. Actualment soc enginyer de #DevOps #Leadtech ;).
Fan de codi obert, usuari de #Linux dels anys 90 i en l'actualitat ambaixador d'#OpenSUSE a Espanya.
Just ara amb l'edat que tinc, m'ha donat pel #Gaming
Algú em recordarà com [^BgTA^] antic #IRCop de #Terra
#Introduction #Introducció #Introductions #AboutMe #Català #Programming #IT #Developer #Technology #TDAH #TEA #NoMoa
-
Apassionat de les noves tecnologies, ment autodidacta i inquieta. Estic al món de la informàtica des dels 12 anys. Actualment soc enginyer de #DevOps #Leadtech ;).
Fan de codi obert, usuari de #Linux dels anys 90 i en l'actualitat ambaixador d'#OpenSUSE a Espanya.
Just ara amb l'edat que tinc, m'ha donat pel #Gaming
Algú em recordarà com [^BgTA^] antic #IRCop de #Terra
#Introduction #Introducció #Introductions #AboutMe #Català #Programming #IT #Developer #Technology #TDAH #TEA #NoMoa
-
Apassionat de les noves tecnologies, ment autodidacta i inquieta. Estic al món de la informàtica des dels 12 anys. Actualment soc enginyer de #DevOps #Leadtech ;).
Fan de codi obert, usuari de #Linux dels anys 90 i en l'actualitat ambaixador d'#OpenSUSE a Espanya.
Just ara amb l'edat que tinc, m'ha donat pel #Gaming
Algú em recordarà com [^BgTA^] antic #IRCop de #Terra
#Introduction #Introducció #Introductions #AboutMe #Català #Programming #IT #Developer #Technology #TDAH #TEA #NoMoa
-
Apassionat de les noves tecnologies, ment autodidacta i inquieta. Estic al món de la informàtica des dels 12 anys. Actualment soc enginyer de #DevOps #Leadtech ;).
Fan de codi obert, usuari de #Linux dels anys 90 i en l'actualitat ambaixador d'#OpenSUSE a Espanya.
Just ara amb l'edat que tinc, m'ha donat pel #Gaming
Algú em recordarà com [^BgTA^] antic #IRCop de #Terra
#Introduction #Introducció #Introductions #AboutMe #Català #Programming #IT #Developer #Technology #TDAH #TEA #NoMoa
-
Apassionat de les noves tecnologies, ment autodidacta i inquieta. Estic al món de la informàtica des dels 12 anys. Actualment soc enginyer de #DevOps #Leadtech ;).
Fan de codi obert, usuari de #Linux dels anys 90 i en l'actualitat ambaixador d'#OpenSUSE a Espanya.
Just ara amb l'edat que tinc, m'ha donat pel #Gaming
Algú em recordarà com [^BgTA^] antic #IRCop de #Terra
#Introduction #Introducció #Introductions #AboutMe #Català #Programming #IT #Developer #Technology #TDAH #TEA #NoMoa
-
Deixar de fer servir #C i #C++ i centrar-se en #Rust i #Java: això recomana l'Agència de Seguretat dels #EUA a empreses del món.
-
Avui a #Leadtech hem tingut un meet sobre #Mastodon i #ActivityPub amb la fantàstica gent de #SocialMedia
Es un primer pas al #Fediverse
Gràcies @andreina_p per la paciència 😜
Com diem a #Catalunya : Mica en mica s'omple la pica
-
S'ha de repetir! 🤗
RT @leadtech_social
🎉Después de 2 años y medio, pudieron ponerse “𝙋𝙞𝙚𝙧𝙣𝙖𝙨 𝙮 𝙥𝙞𝙚𝙨” todos los compañeros y compañeras que se conocían sólo a través de la pantalla… Así fue la gran celebración de Leadtech con la fiesta 🎉“𝗟𝗲𝗴𝘀 & 𝗙𝗲𝗲𝘁”
#leadtech #leadteam #barcelona
https://twitter.com/leadtech_social/status/1582384667184877568