VibeVoice-ASR-Streaming-1.5B vs Granite Speech 5.0 470M TurboCTC: the two quiet open-weights streaming bets

The two most interesting open-weights streaming speech-to-text releases of late August 2026 both shipped without a launch event. On August 25, IBM put Granite Speech 5.0 470M TurboCTC on Hugging Face — weights, model card, and a blog post, nothing more — and on September 2, Microsoft Research uploaded VibeVoice-ASR-Streaming-1.5B under the microsoft namespace with no announcement at all beyond a news line in its VibeVoice GitHub repository. They are the same shape of thing and opposite engineering bets: IBM's model is a 470-million-parameter pure-CTC encoder designed to run at edge speeds on a laptop, while Microsoft's is a 1.5B-class speech language model wrapped in audio tokenizers and a diffusion head — about three billion parameters in total — built to understand speech as language while it transcribes.

Before the numbers, the sourcing labels that keep this honest. Everything marked IBM-reported comes from the Granite Speech 5.0 470M TurboCTC model card and its Open ASR leaderboard entries, run through the official harness by IBM on release day and not yet reproduced independently. Everything about VibeVoice-ASR-Streaming-1.5B comes from reading the repository — the config files, the model card, and Microsoft's streaming docs — because Microsoft has published no accuracy or latency numbers in text and nobody outside Microsoft has benchmarked the checkpoint. The two were not run in a shared harness; this comparison reads both against the same public evidence, which is all either company has offered so far.

Two quiet releases, eight days apart

Both models arrived in the same unceremonious way, which is worth stating plainly because neither company has said much since. IBM's Granite Speech 5.0 470M TurboCTC is the Apache-2.0 member of a two-variant launch — IBM also published a non-commercial -NC sibling with slightly better headline numbers — and its card carries a release date of August 25, 2026, with native Transformers support and an Open ASR leaderboard submission dated the same day. Microsoft's streaming pair, VibeVoice-ASR-Streaming-7B and VibeVoice-ASR-Streaming-1.5B, were created on Hugging Face within the same minute on September 2; the GitHub news line announcing "a unified streaming ASR model that continuously transcribes who said what as speech arrives" is dated September 3 and names the 7B, leaving the 1.5B covered here as the smaller sibling that shipped quietly beside it.

The interesting part is that two teams reached for the word "streaming" and built opposite things. Granite Speech 5.0 470M TurboCTC is a conformer encoder trained with CTC and decoded in one non-autoregressive pass — there is no decoder generating text token by token, so the model is structurally cheap and structurally fast. VibeVoice-ASR-Streaming-1.5B is a speech LLM: two convolutional tokenizers turn 24 kHz audio into a ~7.5 Hz speech-token stream, a Qwen2-family decoder (28 layers, 1,536 hidden units — the 1.5B class) reads it, and a diffusion head produces the transcript in chunks of about 2.9 seconds with roughly half a second of lookahead. One is a race car; the other is a small language model that happens to listen.

The spec check

• Parameters — Granite Speech 5.0 470M TurboCTC: 470M, encoder-only vs VibeVoice-ASR-Streaming-1.5B: ~3B total (1.5B-class LM backbone plus tokenizers and diffusion head).

• Architecture — conformer encoder with block self-attention and self-conditioning, CTC-trained, greedy non-autoregressive decode vs dual acoustic/semantic tokenizers feeding a Qwen2-family causal decoder with a DDPM diffusion head.

• Output cadence — ~12.5 output frames per second, streaming chunkwise vs ~7.5 Hz speech tokens, emitting text per ~2.9 s chunk with ~0.5 s lookahead.

• Languages — English only vs ten: Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.

• Weights — roughly 0.5B params / a fraction of a GB, edge-class vs ~5.6 GB bf16, NVIDIA-GPU-class.

• Accuracy in print — IBM-reported ~5.00% average WER on the Open ASR short-form sets vs no WER figure published in text.

• License — Apache-2.0 vs MIT.

• Released — August 25, 2026 vs September 2, 2026.

Speed: one is built to be small, the other to be a language model

The architectural divide shows up first in speed and hardware. Granite Speech 5.0 470M TurboCTC is pitched at laptops, smartphones, and other edge devices. Because a pure CTC encoder samples nothing — output is a single greedy pass over a 16,384-unit BPE head — it is extremely cheap to run, and IBM's model card reports an Open ASR throughput figure measured on a single H200 that is in the tens of thousands of RTFx for the TurboCTC line, alongside a WebGPU demo that transcribes live in a browser tab. Those figures are IBM-measured and unreproduced, but the structural point stands on its own: a 470M encoder with no autoregressive decoder is a model you can run on hardware a phone ships with.

VibeVoice-ASR-Streaming-1.5B makes the opposite trade. Its decoder is a causal language model, which means text is generated in a heavier loop than a CTC argmax, and the documented ways to run it are self-hosted on NVIDIA GPUs — the microsoft/VibeVoice repository's Python demos or its vLLM plugin — with ~5.6 GB of bf16 weights before any KV cache. Microsoft has published no throughput or real-time-factor claim, so there is no vendor number to set against IBM's. What the LLM architecture buys instead is context: the model carries speaker-and-content understanding across the transcript the way a language model does, which is the difference between transcribing sound and following a conversation.

Accuracy: one vendor printed numbers, the other printed a picture

On evidence, Granite Speech 5.0 470M TurboCTC is ahead of VibeVoice-ASR-Streaming-1.5B by an entire publishable benchmark. IBM ran its model through the official Open ASR leaderboard harness on release day and reports an average word error rate around 5.00% on the public English short-form sets — a figure that is vendor-measured, unverified by any third party, and entirely absent from Microsoft's side of this comparison. VibeVoice-ASR-Streaming-1.5B's model card ships an evaluation results figure as an image but no numeric WER, RTF, or latency in prose; the only numeric anchor in the VibeVoice family is the batch VibeVoice-ASR card's vendor-reported 7.77% average WER across eight English test sets, which describes the non-streaming model and does not transfer. Whichever way the eventual independent runs land, the honest summary today is that IBM published a number and Microsoft published a picture.

Languages and output: English-only versus who-said-what in ten

Granite Speech 5.0 470M TurboCTC is English-only and returns raw transcribed text. That is not a shortcoming for the workloads IBM is aiming at — enterprise English transcription on edge hardware — but it ends the comparison the moment your audio is not English. VibeVoice-ASR-Streaming-1.5B lists ten languages and claims streaming speaker-attributed output: the model card says it "continuously transcribes who said what as speech arrives," with hotwords for domain terms passed as a context prompt. Both extras deserve a skeptical reading — streaming diarization across chunk boundaries is genuinely hard, and the claim is unverified — but they are the reason a team with live multilingual meetings would look at Microsoft's model at all.

Licensing and ecosystem: Apache-2.0 versus MIT, Transformers versus a research repo

Both licenses permit commercial use, which already separates this pair from IBM's own -NC variant of the same architecture. Granite Speech 5.0 470M TurboCTC is Apache-2.0 with the usual patent grant, and it is native in Hugging Face Transformers (AutoModelForCTC from transformers >= 5.16) plus mlx-audio for Apple Silicon — meaning it runs anywhere the largest deployment community in the ecosystem runs, on any vendor's hardware. VibeVoice-ASR-Streaming-1.5B is MIT, which is simpler still, but its serving path is a from-source research setup: the checkpoint's architecture class, VibeVoiceForASRStreamingTraining, is not in the public Transformers documentation, and Microsoft's own streaming docs point to the repo's demo code and a vLLM plugin rather than a stock library load. For a production team, the difference is real: one model drops into an existing Transformers pipeline today, and the other asks you to stand up a research repo and verify the load path yourself.

Which quiet release should you evaluate?

Evaluate Granite Speech 5.0 470M TurboCTC first if your workload is English, your hardware is edge-class or CPU-bound, and you want a model that drops into Transformers with a number on the card to sanity-check against. It is the lower-risk bet in every dimension except one — it will not help you with a second language or a live meeting transcript that needs to know who is speaking.

Evaluate VibeVoice-ASR-Streaming-1.5B if you need multilingual streaming, if who-said-what output or hotwords are features you want to test, or if you would rather bet on a language-model approach to speech than a pure encoder. Budget for an NVIDIA GPU, a from-source install, ~5.6 GB of weights, and an evaluation you design yourself, because nobody — Microsoft included — has published a number you can rely on yet.

What neither quiet release changes is the value of keeping the ASR layer swappable while you find out which bet pays. Both models are young and neither is served through OrcaRouter as of this writing; when a provider hosts either one, the low-risk way to trial it is behind a routing layer that fronts 200+ models at provider list price — 0% markup, so price changes land same-day — with automatic failover to whatever you already trust. Route a slice of real audio to the new model, read the transcripts, and let your own traffic, not a release-day benchmark sheet, decide which quiet bet was the right one.