VibeVoice-ASR-Streaming-1.5B, Explained: Microsoft's Streaming ASR Arrived Without a Launch

On September 2, 2026, Microsoft Research uploaded two new speech-to-text checkpoints to Hugging Face with no press release, no blog post, and no fanfare: microsoft/VibeVoice-ASR-Streaming-1.5B and its larger sibling microsoft/VibeVoice-ASR-Streaming-7B. The only announcement so far is a news entry dated September 3, 2026 in the News section of Microsoft's open-source VibeVoice GitHub repository, describing the release as a unified streaming ASR model that transcribes who said what while the audio is still arriving, with customized hotwords and ten languages. Both checkpoints are MIT-licensed, both were created within about five minutes of each other on the official microsoft Hugging Face account, and both showed zero downloads when we checked a day later. We could find no third-party coverage of either one.

That makes this an unusual write-up: a real, datable release — weights on Hugging Face dated September 2, an in-repo announcement dated September 3 — that the wider world has not caught up to yet. Everything below that is knowable comes from reading the repository: the config files, the model card, and Microsoft's own streaming docs. Everything that is not yet confirmed — accuracy, real latency, whether the streaming speaker attribution survives a real two-speaker call — is labeled as such. There is no benchmark to cite because Microsoft has not published one in text form yet.

The only announcement is a news line

VibeVoice is Microsoft Research's family of MIT-licensed, open-source speech models. Its best-known ASR member, microsoft/VibeVoice-ASR, shipped on January 21, 2026: a batch model that ingests up to sixty minutes of audio in a single pass and returns structured transcripts with speaker, timing, and content — roughly 700,000 Hugging Face downloads have followed. On September 2 the same organization published the streaming siblings, microsoft/VibeVoice-ASR-Streaming-7B and microsoft/VibeVoice-ASR-Streaming-1.5B; Hugging Face's API timestamps the 7B at 2026-09-02T15:46 UTC and the 1.5B five minutes later at 15:51 UTC. The GitHub repository's model table now lists VibeVoice-ASR-Streaming as its own entry, and its News section carries the September 3 line announcing it.

What is genuinely new versus the January model is the interaction model: the streaming checkpoints transcribe as audio arrives instead of waiting for a complete file. Everything else — the underlying speech-token architecture, the speaker-attributed output, the hotword mechanism — is a continuation of the batch design, scaled and re-pointed at real-time use. Note the language difference up front: the batch model advertises 50+ languages, while the streaming card lists ten.

What "streaming" means in this checkpoint

Microsoft's streaming doc describes the intent plainly: the model transcribes while the audio is still arriving, and it emits text once per audio chunk, so a transcript appears as the speaker talks. The 1.5B repo's preprocessor_config.json makes the cadence concrete:

• Sample rate and token rate — 24 kHz audio in, compressed 3,200×, which yields roughly a 7.5 Hz stream of speech tokens (about one token every 133 ms).

• Chunk — 22 frames, which at 7.5 Hz is about 2.9 seconds of audio per emitted segment.

• Lookahead — 4 frames, about 0.5 seconds of future audio used to firm up the current segment.

(The arithmetic is straightforward from the repo: 22 × 3,200 = 70,400 samples ≈ 2.93 s at 24 kHz; 4 × 3,200 = 12,800 samples ≈ 0.53 s. Microsoft's own doc notes that a checkpoint always runs at the chunk it was trained on, so these are not tunable at inference time.)

That places this in the chunked-streaming family rather than the word-by-word one: a live transcript grows in roughly three-second increments with about half a second of lookahead, not in per-token partials. That is a legitimate and common design for meeting and call transcription, but it is a different latency profile from systems that emit sub-second partials — so treat "streaming" as a spectrum and measure it against your own latency budget before building a live-captioning product on it.

Under the hood: a Qwen-scale speech LLM

Reading the 1.5B checkpoint's config.json gives the now-familiar VibeVoice recipe at a smaller scale:

• The decoder is a Qwen2-family language model whose dimensions match the 1.5B Qwen2.5 class exactly — 28 layers, 1,536 hidden units, 12 attention heads with 2 key-value heads, and a 65,536-token window. The LLM is what lets the model carry speaker and content understanding across the transcript.

• Acoustic and semantic tokenizers — deep convolutional encoders with 8/5/5/4/2/2 striding — turn 24 kHz audio into the ~7.5 Hz speech-token stream the decoder reads.

• A diffusion head (DDPM, 20 denoising steps, v-prediction) sits on the generation side, consistent with the rest of the VibeVoice line.

• Size honesty: the checkpoint's own safetensors metadata lists about 3 billion parameters, roughly 5.6 GB of weights. The "1.5B" in the name is the scale of the language backbone — the natural reading of a config whose decoder is a 1.5B-class Qwen model — with the audio tokenizers and diffusion head adding the rest. The architecture string in the config, VibeVoiceForASRStreamingTraining, signals that this is a research-family checkpoint.

Who said what, in ten languages

The model card's headline capability is streaming speaker-attributed transcription: it "continuously transcribes who said what as speech arrives," per the card's own bullet. It also advertises customized hotwords for domain terms, and lists ten supported languages — Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.

Hotwords are implemented through the same context-biasing mechanism the batch model uses. In Microsoft's command-line demo you pass them as context information — for example --context_info "Microsoft,VibeVoice" — to bias recognition toward names and technical terms without any fine-tuning. The card says nothing about language auto-detection or code-switching for the streaming model, which is another gap versus the batch model's claims.

One caveat deserves emphasis. The streaming documentation page concentrates on chunked emission and hotwords; it does not detail how speaker attribution is maintained across chunk boundaries. Streaming diarization is genuinely hard — speakers overlap, and a chunk boundary is exactly where attribution drifts. The "who said what" framing on the card is a vendor claim until someone runs a real two-speaker live call through it and checks.

Where it sits: the VibeVoice family and the 2026 streaming-ASR field

Inside the family, the release slots in as follows:

• microsoft/VibeVoice-ASR (January 21, 2026) — the batch flagship: up to 60 minutes in one pass, 50+ languages, Who/When/What output, hotwords, roughly 700K downloads.

• microsoft/VibeVoice-ASR-Streaming-7B and microsoft/VibeVoice-ASR-Streaming-1.5B (September 2, 2026) — the new streaming variants; this piece's subject is the 1.5B.

• microsoft/VibeVoice-ASR-BitNet (July 23, 2026) — a quantized, CPU-oriented edge engine for the batch model.

• The VibeVoice-1.5B TTS and VibeVoice-Realtime-0.5B streaming-TTS models are separate members of the same family, not part of the ASR line.

The wider open-weights ASR field has been moving toward real-time all year, so a new streaming entry has direct points of comparison. Qwen3-ASR (Alibaba, ~1.7B) is a unified streaming-and-offline model covering 50+ languages and dialects. NVIDIA's Nemotron 3.5 ASR is a 0.6B streaming model spanning 40 languages, under the OpenMDW license rather than MIT. IBM's Granite Speech 5.0 470M TurboCTC is Apache-2.0 with vendor-reported throughput figures that are unusually high. Against those, Microsoft's streaming model differentiates on three things: MIT licensing, an LLM backbone that carries speaker-and-content understanding rather than a pure acoustic model, and the speaker-attributed live-transcription framing. Its open questions are accuracy and real-world latency — neither has an independent number yet.

What is not yet verified

Reading the repo tells you the design; it does not tell you how well it works. Specifically:

• No accuracy or latency numbers in text. The model card ships a results figure as an image, but no numeric WER, RTF, or latency table in prose — nothing to independently cite.

• No third-party evaluation. Downloads were at zero a day after upload; there is no community benchmark, no leaderboard entry, and no independent test we could find.

• The serving path is a from-source research setup. Microsoft's streaming docs run from a clone of the VibeVoice GitHub repo inside an NVIDIA PyTorch container (nvcr.io/nvidia/pytorch, with flash-attention recommended). The model card also shows a Transformers pipeline snippet and defers installation details to GitHub; whether the current released Transformers version runs streaming inference on this checkpoint out of the box, we have not verified.

• The framing is research-first. Microsoft's repository describes the VibeVoice models as intended for research and development purposes. The weights are MIT; the posture is "here is the science," not "here is a supported product." Budget for your own evaluation gate.

Should you build on it?

For teams that already self-host ASR, this is worth a weekend evaluation if your audio is inside the ten-language set and you specifically need speaker-attributed live transcription from an MIT-licensed model. Budget for ~5.6 GB of weights for the 1.5B on an NVIDIA GPU and a from-source install, and plan to measure accuracy yourself on your own audio before trusting it.

For teams shipping a production transcription pipeline today, the prudent answer is to wait on one of three things: Microsoft publishing the accuracy and latency figures that currently exist only as an image; an independent benchmark; or a maintained serving path with a supported runtime. The chunked ~3-second cadence is also a spec to sanity-check against your latency requirement rather than assume from the word "streaming."

The cheap way to keep the option open is to avoid hardwiring your application to a single transcription engine. A routing layer that fronts many models through one API means that when VibeVoice-ASR-Streaming — or the next open ASR — lands on an inference provider, A/B-testing it against your incumbent on the same traffic is a configuration change rather than a re-platforming. Because a pass-through router charges the provider's list price with no markup, that comparison stays cheap, and automatic failover keeps a young, unproven model from becoming a single point of failure in your pipeline. That is the general pattern for adopting any days-old model: let it earn a place on real traffic before you bet production on it.

FAQ

Is VibeVoice-ASR-Streaming-1.5B an actual Microsoft release?

Yes. The checkpoint sits on the official microsoft Hugging Face account with a creation timestamp of September 2, 2026, and the microsoft/VibeVoice GitHub repository references VibeVoice-ASR-Streaming in its model table with a news entry dated September 3, 2026. What is unusual is not the provenance but the silence: no press release, no blog post, and no third-party coverage as of this writing.

Why two sizes, 7B and 1.5B?

Microsoft uploaded both checkpoints in the same minute but has not published a comparison. From the configs, the 1.5B is the small end — a 1.5B-class Qwen language backbone plus the audio stack, about 3B parameters and ~5.6 GB of weights. The natural reading is a higher-accuracy 7B and a cheaper, faster 1.5B, but Microsoft has not said so, and there are no benchmarks to confirm the trade-off.

How is this different from the earlier VibeVoice-ASR?

The January batch model ingests up to 60 minutes of audio in a single pass and advertises 50+ languages. The streaming checkpoints transcribe while the audio is arriving, emitting a transcript in ~3-second chunks with ~0.5 seconds of lookahead, and the model card lists ten languages. Both share the same speech-token architecture, the speaker-attributed output idea, and the hotword mechanism.

For now, VibeVoice-ASR-Streaming-1.5B is a checkpoint plus a news line. Read the repo if you self-host and need a permissively licensed streaming ASR to evaluate; hold for the numbers if you are choosing a production engine. The interesting signal is that Microsoft is pushing its voice line toward real time at two sizes at once — and did it so quietly that a day later, the downloads counter still read zero.