Interfaze, a young YC’s startup, has open-sourced a new speech recognition model. It is called diffusion-gemma-asr-small
. The model transcribes audio through a diffusion decoder, not an autoregressive one. It is described as the first multilingual audio diffusion ASR model. One adapter handles six languages. The research team trained only about 42M parameters on top of a frozen 26B backbone. That is roughly 0.16% of the model’s weights.
Here two terms matter up front. Autoregressive models generate text one token at a time. Diffusion models refine all tokens in parallel. This model uses the diffusion approach for speech-to-text.
TL;DR
- Claimed by the Interfaze team, to be the first open-source multilingual diffusion ASR: six languages from a single ~42M-parameter adapter.
- Transcribes via DiffusionGemma’s diffusion decoder using uniform, random-token diffusion, not the absorbing
<mask>
scheme. - Transcription cost scales with denoising steps, not transcript length.
- Leads diffusion peers on LibriSpeech (6.6% WER vs Whisfusion’s 8.3%) but trails autoregressive Whisper.
- The adapter ships under Apache-2.0; DiffusionGemma (Gemma terms) and whisper-small (MIT) load separately.
What is diffusion-gemma-asr-small?
diffusion-gemma-asr-small
is an audio-native ASR model. It converts speech to text using a discrete diffusion decoder. That decoder belongs to DiffusionGemma, Google’s 26B mixture-of-experts model. DiffusionGemma activates 4B parameters, using 128 experts with top-8 routing. It generates text by discrete diffusion instead of autoregression.
The diffusion detail is specific. Most diffusion LLMs use an absorbing <mask>
scheme. DiffusionGemma uses uniform, random-token diffusion instead. It fills a fixed-length canvas with random vocabulary tokens. Each step keeps confident predictions and re-randomizes the rest. After a few steps the noise anneals into text.
Interfaze added audio to this text-only model. Out of the box, DiffusionGemma takes text, images, and video. It does not take audio. The repo ships only the trained adapter, about 42M parameters. The frozen backbones download separately from their own repos.
How it works
The model does not feed raw waveforms to the LLM. An early attempt tried exactly that and failed. A frozen LLM has never seen a spectrogram. The embedding space has no notion of formants or phonemes. The model learned to ignore audio and hallucinate fluent nonsense.
The working design uses a frozen whisper-small
encoder. It acts only as a feature extractor, not a decoder. Whisper turns 30 seconds of audio into 1500 frames. Each frame holds 768-dimensional acoustic features. A small trainable projector then compresses these frames. It uses conv layers that subsample 8× plus a linear map. The output is 188 “audio tokens” at 2816 dimensions. These tokens scatter into the prompt’s reserved <|audio|>
slots. LoRA adapters let the backbone attend to this new modality. The decoder then denoises a 192-token transcript canvas. It runs bidirectionally over roughly 16 steps.
The pipeline, from the model card, is compact:
raw audio ─► whisper-small encoder (frozen) ─► projector (trained, ~19M)
─► scatter into <audio> token slots of DiffusionGemma's encoder
─► DiffusionGemma decoder denoises a 192-token canvas (bidirectional, cross-attends audio)
─► transcript
The training unlock
The first training runs stalled. Loss flatlined near 8. The failure was circular. The projector started random, so its output was noise. Attention then learned to ignore it. Almost no gradient reached the projector. The model never learned.
The fix supervised the projector directly. The research team ran the 188 audio tokens through DiffusionGemma’s frozen lm_head
. They applied a CTC loss against the transcript. CTC means Connectionist Temporal Classification. It aligns audio features to text without needing attention.
This sidesteps the standoff. The audio embeddings became linearly predictive of the right words. CTC loss then dropped from 24 to 8.6 in 300 steps. On LibriSpeech test-clean, English WER fell 90% → 52% → 14.6% → 6.6% over ten epochs.
Performance and benchmarks
WER means Word Error Rate, where lower is better. CER means Character Error Rate. The model trained on FLEURS, LibriSpeech, and VoxPopuli. All scores below use the Whisper text normalizer at 16 diffusion steps.
| benchmark | metric | score |
|---|---|---|
| LibriSpeech test-clean (en) | WER | 6.6% |
| FLEURS English | WER | 15.7% |
| VoxPopuli English | WER | 18.5% |
| FLEURS Hindi | CER | 15.8% |
| FLEURS Mandarin | CER | 29.6% |
Against other diffusion or non-autoregressive ASR, it leads.
| model | approach | LibriSpeech test-clean |
|---|---|---|
| TransFusion (2022) | multinomial diffusion | ~6–7% (proof-of-concept) |
| Whisfusion (Aug 2025) | Whisper-large-v3 + masked diffusion | 8.3% |
| diffusion-gemma-asr-small (2026) | Whisper-small + DiffusionGemma | 6.6% |
Against autoregressive Whisper, it trails. The team frames this gap as data, not architecture.
| benchmark | ours | Whisper-small | Whisper-large-v3 |
|---|---|---|---|
| LibriSpeech clean | 6.6% | ~3.4% | ~2.0% |
| FLEURS-en | 15.7% | ~9–10% | ~4–5% |
| VoxPopuli-en | 18.5% | ~9–11% | ~7–10% |
The denoising-step sweep shows a nearly flat curve.
| steps | FLEURS-en WER | speed |
|---|---|---|
| 8 | 15.7% | 14.9× real-time |
| 16 | 15.6% | 10.3× |
| 32 | 15.2% | 6.5× |
| 48 | 15.6% | 4.7× |
Going from 8 to 48 steps buys about 0.1 WER point. It costs roughly 3× the latency. The model converges in about 8 parallel passes. That is around 0.7–1.5s of model time for a 10-second clip.
Use cases with examples
- Batch transcription pipelines benefit from parallel decoding. Cost is set by denoising steps, not clip length. A 10-second clip needs roughly the same passes as a shorter one.
- Multilingual transcription runs from a single adapter. It covers English, German, French, Spanish, Hindi, and Mandarin. Teams avoid a separate model per language.
- Non-autoregressive ASR research gains a reproducible baseline. The recipe grounds a frozen LLM with a small adapter. Researchers can extend it with more audio or a larger encoder.
How to get started
The model lives on the Hub. It ships the adapter, model.py
, audio.py
, and a runnable inference.py
. DiffusionGemma support needs transformers
from main.
pip install torch peft soundfile librosa huggingface_hub \
"transformers @ git+https://github.com/huggingface/transformers.git"
Then transcribe in Python:
import sys, soundfile as sf
from huggingface_hub import snapshot_download
repo = snapshot_download("interfaze-ai/diffusion-gemma-asr-small") # adapter, ~170 MB
sys.path.insert(0, repo)
from inference import load, transcribe
model, tok, fe = load(f"{repo}/diffusion_asr_small.pt", device="cuda")
wav, sr = sf.read("audio.wav") # 16 kHz mono float32
print(transcribe(wav, model, tok, fe, max_steps=16))
A command-line path also works from inside the downloaded repo:
python inference.py audio.wav
The max_steps
argument trades speed for accuracy. The team notes 8 is near-best and fastest. The default is 16. The base models load under their own licenses: DiffusionGemma under Gemma terms, whisper-small under MIT.
Interactive Explainer
Check out the** Model weights **and
.
Technical detailsAlso, feel free to follow us on
and don’t forget to join ourTwitter
and Subscribe to
150k+ML SubReddit. Wait! are you on telegram?
now you can join us on telegram as well.Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.