On October 21, IBM updated its Granite Speech family on Hugging Face in one move: a new flagship, granite-speech-5.0-2b, plus revised versions of two models the community already runs — the edge-oriented granite-speech-5.0-470m-turboctc and the meeting-focused granite-speech-4.1-2b-plus. Updating three models at once is unusual in the speech track, and it signals an attempt to cover the three core enterprise speech needs in a single sweep.

What each model handles

The 2B flagship handles accuracy. granite-speech-5.0-2b succeeds granite-speech-4.1-2b at the same parameter count. It transcribes English, French, German, Spanish, Portuguese, and Japanese, translates to and from English in those languages, and adds English-to-Italian and English-to-Mandarin directions. Training used 390,000 hours of public audio plus synthetic data built for Japanese ASR, keyword biasing, and translation. It supports keyword biasing lists of up to 2,000 terms — feeding product names, people, and acronyms directly to the model — and punctuation and capitalization on request, with capitalization F1 ranging from 89.7 to 99.5 on IBM's test sets.

The 470M TurboCTC handles speed. Launched in August as an English-only 470M encoder, this revision expands to seven languages (English, Japanese, Spanish, German, French, Italian, Portuguese), outputs native punctuation and capitalization, and adds training that targets reverberant audio. The speedup comes from compiling a single Conformer block and reusing its graph across the stack: RTFx on the Open ASR Leaderboard rose from roughly 12,600 at launch to 19,870 — about 1.6x faster at the same accuracy.

The 4.1-2b-plus handles "who spoke." Architecture, prompt format, and API are unchanged, making it a drop-in weight swap. Official numbers: word diarization error rate (WDER) dropped from 4.40 to 4.10, a 7% cut; speaker-attributed WER (cpWER) fell from 30.50 to 28.48; long-form WER dropped 24% (13.53 to 10.22); biased-keyword errors fell 45% with 20-term lists (11.79 to 6.48); and average timestamp shift on all speech was cut from 210.86 to 98.54 — more than half, a direct win for captions and search.

Three observations beyond the numbers

First, the enterprise speech stack is stratifying. Accuracy, edge speed, and speaker attribution are three different needs, and IBM covers them with three purpose-built models rather than one "do-everything" model. Set against the same-week Qwen-Audio-3.1-ASR and Voxtral 3 releases, open speech supply is getting crowded fast.

Second, licensing and inference stacks are fully covered. All three ship under Apache 2.0 with native Transformers support; the 2B flagship also runs on vLLM, llama.cpp, and mlx-audio for Apple Silicon, and TurboCTC has a browser WebGPU streaming demo. For teams avoiding cloud API lock-in, this is a complete self-hosted path.

Third, read the leaderboard numbers with attribution in mind. The "5.06% average English WER, No.2 among open-weight models" figure is IBM's own reporting on the Open ASR Leaderboard's public English short-form sets as of October 21; the 1.6x speedup is likewise an official claim, and independent retests have not yet appeared.

Reference: IBM Granite official blog https://huggingface.co/blog/ibm-granite/granite-speech-5

For enterprise users the signal is clear: speech is replaying the script LLMs already ran — open-weight family releases, clear division of labor, edge-deployable. The next question is who first turns "transcribe + attribute + translate" into a genuinely smooth end-to-end product.