VibeVoice-ASR-Streaming-1.5B vs Gemini 3.5 Transcribe: Microsoft's quiet streaming ASR against Google's new API

Seven days and one philosophical divide separate the two newest streaming speech-to-text systems. On August 26, 2026 Google put Gemini 3.5 Transcribe into public preview: a hosted, closed speech-to-text API priced per audio token, with a separate Live endpoint for real-time sessions. On September 2, 2026 Microsoft Research uploaded VibeVoice-ASR-Streaming-1.5B to Hugging Face under the microsoft namespace — MIT-licensed weights, no press release, no launch post, and at this writing no third-party coverage anywhere. Both transcribe speech while it is still arriving. Almost everything else about them — who is allowed to run them, what they cost, what evidence exists that they work — is different.

This page compares the two as they actually exist today, which means the two sides of the evidence are not symmetric. Gemini 3.5 Transcribe is a live Google product with published token prices and third-party accuracy figures from Artificial Analysis; those are the numbers labeled AA below. VibeVoice-ASR-Streaming-1.5B is a checkpoint whose model card publishes no word-error-rate and no latency figure in text at all — the card's evaluation is an image, and the streaming family's only announcement so far is a news line dated September 3 in Microsoft's VibeVoice GitHub repository. Everything on the Microsoft side below is what is knowable from the repository: the config files, the model card, and Microsoft's own streaming documentation. Nothing about it has been independently benchmarked yet.

The two models at a glance

• What it is — VibeVoice-ASR-Streaming-1.5B: open checkpoint, MIT license, self-hosted vs Gemini 3.5 Transcribe: hosted API, closed weights.

• Release — Sept 2, 2026 weights, Sept 3 repo news line vs Aug 26, 2026 public preview with full launch materials.

• Streaming shape — emits text in ~2.9-second chunks with ~0.5 s lookahead, session state carried in a prefix cache vs WebSocket Live session billed per audio token, capped at 10 minutes of audio per session.

• Accuracy in print — no WER or latency figure published as text vs AA-measured 2.6% WER on the pre-recorded endpoint and 4.0% on the Live endpoint.

• Speaker output — claims streaming who-said-what attribution (unverified) vs no speaker diarization and no word-level timestamps on the Live endpoint.

• Hotwords — supported, via the same context prompt as the batch VibeVoice-ASR vs not a listed feature of the Live endpoint.

• Cost — $0 for the weights, ~5.6 GB to self-host vs roughly $0.30 per audio-hour blended on the pre-recorded endpoint and ~$0.54 on Live, per Google's listed token prices.

A hosted API and a quiet upload, seven days apart

Gemini 3.5 Transcribe is Google's replacement-grade transcription model, and it ships as two products. The pre-recorded endpoint, served through the Interactions API, takes a full audio file and returns a transcript with the expected extras. The Live endpoint, gemini-3.5-transcribe-live, runs over a bidirectional WebSocket and is the one that competes with VibeVoice-ASR-Streaming-1.5B on the shape of the job: transcribing a session as it happens. Google announced it with the normal machinery — a public-preview launch on August 26, pricing on the model page, third-party leaderboards picking it up within days.

Microsoft's streaming release had none of that. On September 2 the microsoft Hugging Face account created two checkpoints in the same minute: VibeVoice-ASR-Streaming-7B and VibeVoice-ASR-Streaming-1.5B. The GitHub repository's model table now lists VibeVoice-ASR-Streaming as a family member, and its News section carries the September 3 line announcing "a unified streaming ASR model that continuously transcribes who said what as speech arrives, with support for customized hotwords and 10 languages" — but that announcement names the 7B, and the 1.5B is the smaller sibling that shipped alongside it without being called out. When we checked a day after upload, both checkpoints still showed zero downloads.

What "streaming" means on each side

The word streaming hides a real design difference. Microsoft's preprocessor config makes the VibeVoice cadence concrete: audio arrives at 24 kHz and is compressed 3,200× into a stream of speech tokens at about 7.5 Hz (one token every ~133 ms). The config then declares a chunk of 22 frames and a lookahead of 4 frames — about 2.9 seconds of audio per chunk, with roughly half a second of future audio used to firm up the current segment. The model emits text once per resolved chunk, and Microsoft's docs note that context from earlier chunks is preserved through the KV cache, so a long session does not recompute from scratch. The arithmetic is in the config; the practical consequence is a transcript that grows in roughly three-second increments, not in per-word partials.

Google's Live endpoint is a different streaming shape: a bidirectional WebSocket session in which audio is billed at 25 audio tokens per second, engineered for interactive use, but capped at 10 minutes of audio per session and — on the Live endpoint specifically — without speaker diarization or word-level timestamps. For anyone comparing the two as live-transcription engines, the differences that matter are the session cap (10 minutes on Google's Live side, bounded only by your own GPU and memory on the Microsoft side), the emission cadence (~3-second chunks versus whatever the WebSocket session delivers), and the output extras (speaker attribution and hotwords claimed on the Microsoft card, absent from Google's Live feature list).

Accuracy: one side has numbers, the other has a picture

This is the widest gap in the matchup, and it is a gap in evidence, not necessarily in quality. Artificial Analysis measures Gemini 3.5 Transcribe at a 2.6% word error rate on the pre-recorded endpoint and 4.0% on the Live endpoint — the premium you pay for real-time delivery shows up in the error rate, which is a common and honest trade-off for streaming systems. Those are third-party numbers on a neutral leaderboard, published days after launch.

VibeVoice-ASR-Streaming-1.5B has no comparable figure anywhere. The model card ships an evaluation results figure as an image, but no numeric WER, RTF, or latency in prose — nothing quotable and nothing independently checkable. Microsoft has not put streaming accuracy numbers in text for either the 7B or the 1.5B. The only numeric anchor in the whole family is the batch VibeVoice-ASR model card, which reports a vendor-run average 7.77% WER across eight English test sets and 2.20% on LibriSpeech clean — but that describes the non-streaming model, and streaming models typically trade a little accuracy for the latency win. Until someone runs the streaming checkpoint through a public harness, the honest position is that Google's model is measured and Microsoft's is not.

Cost: per audio-hour versus per GPU-hour

Google sells Gemini 3.5 Transcribe by the token, and the pricing splits cleanly across the two endpoints. The pre-recorded endpoint lists $2 per 1M audio-input tokens and $12 per 1M text-output tokens, which works out to roughly $0.30 per audio-hour blended. The Live endpoint lists $3.50 per 1M audio tokens and $21 per 1M text tokens — about $0.54 per audio-hour blended, or roughly 80% more than pre-recorded, which is the price of real-time delivery. Google bills audio at 25 tokens per second, so silence is only free if your client does not stream it; reconnects, duplicated audio, and logging can all add to the real bill. A free tier exists, with the caveat that free-tier content may be used to improve Google products.

Microsoft's model is priced in GPU-hours instead. The 1.5B checkpoint is roughly 5.6 GB of bf16 weights (a 1.5B-class Qwen language backbone plus the audio tokenizers and diffusion head — the safetensors metadata sums to about 3B parameters), and the documented paths to run it are self-hosted: the Python demos in the microsoft/VibeVoice repository, or the vLLM plugin Microsoft describes for serving OpenAI-compatible and WebSocket endpoints. There is no hosted price yet because no inference provider lists the model yet. The cost question is entirely whether your own GPU time is cheaper than Google's per-hour rate, which for any sustained transcription volume it almost certainly is — and the license question is separate from the cost question, because MIT weights are yours to keep in a way that no API subscription is.

The two are not mutually exclusive in a production stack. The pragmatic pattern for a team that needs live transcription today is to keep the ASR layer behind one endpoint so the choice stays reversible: a routing API that fronts 200+ models at provider list prices — no markup, so a vendor price change is live the same day — with automatic failover, is exactly the layer that lets you run Google's measured API as the default while a slice of real traffic tests VibeVoice-ASR-Streaming-1.5B the moment a provider hosts it. That is OrcaRouter's model, and it is the low-risk way to adopt a checkpoint whose accuracy nobody has verified yet.

Who should pick which

Pick Gemini 3.5 Transcribe if you want a transcription API that works today, with published accuracy, predictable per-hour pricing, and no GPU to babysit — and especially if your audio is not inside the ten languages Microsoft lists, or if you need the pre-recorded endpoint's diarization and word-level timestamps for file transcription. The 10-minute Live session cap is a real constraint to test against your use case before you commit to it for live sessions.

Pick VibeVoice-ASR-Streaming-1.5B if you self-host, if you want the weights and the data control that MIT provides, if you need unbounded session length, or if the who-said-what streaming output and hotwords are features you specifically want to evaluate. Budget for a from-source install, ~5.6 GB of weights on an NVIDIA GPU, and your own evaluation gate — there is no benchmark to lean on, so the accuracy question is yours to answer with your own audio.

The one-week verdict: Google shipped the measured product and Microsoft shipped the interesting bet. Gemini 3.5 Transcribe is the safer default and the one with third-party numbers behind it; VibeVoice-ASR-Streaming-1.5B is the open, self-hostable, entirely unverified alternative. What makes the choice cheap is that it does not have to be permanent — keep the ASR endpoint swappable, and a checkpoint that ships without a single benchmark can earn or lose your traffic on the evidence of your own transcripts.