As of 2026-09-01, IBM released Granite Speech 5.0 Turbo CTC on 2026-08-25: two compact 470M English ASR models. IBM reports over 12,600 RTFx on one NVIDIA H200 — more than 3.5 hours of speech in one second with batched inference. BibiGPT turns a video or podcast link into a timestamped transcript and a summary, without a GPU to wire.
Drop an MP3, WAV, or M4A — BibiGPT returns timestamped text you can copy, without wiring a CTC decoder.
As of 2026-09-01, IBM released Granite Speech 5.0 Turbo CTC on 2026-08-25: two 470M English ASR models (Apache 2.0 and CC-BY-NC-SA-4.0). IBM reports over 12,600 RTFx on one NVIDIA H200 — more than 3.5 hours of speech in one second with batched inference. Last updated 2026-09-01.
IBM Research and the ibm-granite org on Hugging Face released two encoder-only English ASR models on 2026-08-25. They drop the LLM backbone used in earlier Granite Speech models so the stack stays at 470 million parameters.
granite-speech-5.0-470m-turboctc is Apache 2.0. The -nc variant trains on extra data (GigaSpeech, SPGI Speech) and is CC-BY-NC-SA-4.0. Pick commercial freedom or a slightly lower WER — not both in one checkpoint.
IBM reports more than 12,600× real-time throughput on a single NVIDIA H200 with batched inference — enough to transcribe more than 3.5 hours of speech in one second. That is IBM's published figure, not an independent bake-off.
On OpenASR Leaderboard public English test sets as of 2026-08-25, IBM reports 5.00% aggregate WER for the Apache model and 4.85% for the NC model. Word error rate counts substitutions, deletions, and insertions — the number is a score, not a promise about names in your recording.
A fast open ASR checkpoint is useful for people who will self-host. Most people with a lecture to catch up on need the words as notes, not a Conformer stack. That second job is what BibiGPT is for.
These models are built for laptops, phones, and high-throughput transcription boxes. They give up speech translation and keyword biasing that earlier Granite Speech models had. You still have to wrap weights, a decoder, and storage.
Raw text from any ASR model still has to be read. BibiGPT takes a link and returns a timestamped transcript plus an AI summary and searchable notes — so a two-hour recording becomes something you can skim.
Both Turbo CTC models are English ASR. Dialect-heavy Chinese, Japanese, or mixed-language podcasts need a different path. BibiGPT transcribes and summarizes those recordings from a pasted link without you picking a checkpoint.
Headline facts from IBM's 2026-08-25 Granite Speech 5.0 Turbo CTC release.
IBM shipped granite-speech-5.0-470m-turboctc (Apache 2.0) and granite-speech-5.0-470m-turboctc-nc (CC-BY-NC-SA-4.0) on Hugging Face the same day as Granite 4.2 language models.
Unlike Granite Speech 4.1, Turbo CTC is encoder-only CTC. IBM dropped speech translation and keyword biasing to keep the footprint small enough for laptops and phones.
IBM reports more than 12,600× real-time throughput with batched inference on a single NVIDIA H200 — more than 3.5 hours of speech in one second. That is a vendor measurement on public OpenASR sets.
As of 2026-08-25 IBM reports 5.00% aggregate WER for the Apache model and 4.85% for the NC model on OpenASR public English tests. On FFASR, IBM says the pair were the two fastest listed models that day.
IBM published a Chrome/Edge WebGPU streaming demo. Both models are English ASR. Mixed-language or dialect-heavy video still needs a broader transcription workflow.
Where a fast open ASR release matters — and where a link-to-notes product is the actual job.
A team transcribing call archives on their own GPUs can evaluate the Apache 2.0 checkpoint. Then still dump the finished files into BibiGPT so humans get chapters, not a raw token stream.
A 90-minute class recording does not need a Conformer stack. Paste the URL into BibiGPT, export the timestamped transcript, and keep asking questions. That is the product path.
Granite Speech 5.0 Turbo CTC is English and throughput-first. Qwen3-ASR covers 52 languages and dialects. Keep one URL per model family; do not merge them into a single 'open ASR' page.
Transcription and notes workflows that pair with this release.
Drop an MP3, WAV, or M4A and get timestamped text.
Turn a video link or file into a searchable transcript.
The multilingual open ASR counterpart from Alibaba's Qwen team.
Another 2026 open ASR event page — 2B, 14 languages.
Pull a full transcript from a YouTube URL.
Facts on this page come from IBM's own 2026-08-25 posts. Throughput and WER are IBM-reported.
IBM Granite announced two 470M English ASR models on 2026-08-25, reporting over 12,600 RTFx on one H200 and OpenASR public-set WER of 5.00% (Apache 2.0) and 4.85% (NC).
Hugging Face — Extremely Fast and Accurate Transcription with Granite Speech 5.0 Turbo CTC ↗Model card for ibm-granite/granite-speech-5.0-470m-turboctc: Apache 2.0, English, 470M, Conformer CTC, release date 25 August 2026.
Hugging Face model card — granite-speech-5.0-470m-turboctc ↗IBM Research's 2026-08-25 Granite 4.2 post names Granite Speech 5.0 Turbo CTC and Turbo CTC NC as the day's speech models (the URL slug is Granite 4.2, not Speech 5.0).
IBM Research — Granite 4.2 ↗Word error rate is the fraction of words that were substituted, deleted, or inserted compared with a reference transcript. A 5.00% WER means five mistakes per 100 reference words on that test set. It does not tell you whether the mistakes are names, numbers, or filler — which is what listeners actually notice.
RTFx is how many seconds of audio a system transcribes per second of wall-clock time. An RTFx of 12,600 means the reported setup processed audio 12,600 times faster than real time. It is a hardware-and-batching number, not a promise about a laptop microphone.
Why people use BibiGPT to turn videos into text every day.
Trusted by 50,000+ users worldwide
“I paste a link and get clean captions in seconds — it saves me hours of retyping every single week.”
Maya R.
Content Creator · Repurposes short videos
“Exporting the transcript lets me review new words at my own pace instead of pausing the video constantly.”
Daniel K.
Language Learner · Studies with real videos
“Accurate, timestamped text I can quote directly. It has quietly become part of my daily workflow.”
Priya S.
Researcher · Cites public talks
FAQ'S
Ask us anything!
1 Best Bilibili AI video summary tool 2026? Paste a link for a free AI recap, mind map, and transcript-style takeaways on 30+ platforms — no login needed.
2 Install a video-summary skill into DeepSeek Harness in one copy-paste — SKILL.md matches Claude Code. Or skip dsh: paste a Bilibili or YouTube link in the browser and get a timestamped summary.
3 Reverse any video free with no app and no download, right in your browser. Step-by-step for iPhone, Android, and PC, plus how to reverse audio (2026).
Paste a YouTube, Bilibili, or podcast link — or upload a file. BibiGPT returns a timestamped transcript, an AI summary, and notes you can search. Last updated 2026-09-01.
