Model reference · open weights
paraformer-zh-streaming is an open-weight audio or speech model from funasr. paraformer-zh-streaming (BF16) weighs 881 MB; the smallest configuration that runs it is RTX 3060 12 GB.
For self-hosting, compare this build's requirements below with the GPU you plan to rent.
Summary of the funasr/paraformer-zh-streaming model card, 2026-11-01
What it is
| Released by | funasr |
|---|---|
| Released | 2024-02-01 |
| VRAM | 881 MB for the weights |
What it runs on
| Card | One stream | Memory |
|---|---|---|
| RTX 3060 12 GB | fits | 11.6 GB |
| RTX 4060 Ti 16 GB | fits | 15.4 GB |
| RTX 3090 24 GB | fits | 23.4 GB |
| RTX 4090 24 GB | fits | 23.4 GB |
| RTX 5090 32 GB | fits | 31.0 GB |
| L40S 48 GB | fits | 44.0 GB |
| A100 80 GB | fits | 78.2 GB |
| H100 80 GB | fits | 78.1 GB |
| RTX PRO 6000 Blackwell 96 GB | fits | 93.8 GB |
| DGX Spark (GB10) 128 GB unified | fits | 107 GB |
| H200 141 GB | fits | 138 GB |
| B200 180 GB | fits | 176 GB |
From the model card
This model is part of the FunASR ecosystem — one industrial-grade open-source toolkit for ASR · VAD · punctuation · speaker diarization · emotion / event · LLM-ASR. A Star really helps the project (and keeps you updated):
🌟 FunASR · 🌟 SenseVoice · 🌟 Fun-ASR · 🌟 FunClip
Streaming speech recognition — real-time transcription with low latency for Chinese.
Streaming variant of Paraformer that processes audio in chunks, delivering results in real-time as audio arrives.
from funasr import AutoModel
model = AutoModel(
model="funasr/paraformer-zh-streaming",
hub="hf",
device="cuda",
)
# Streaming inference with chunk processing
import sounddevice as sd
import numpy as np
chunk_size = [0, 10, 5] # [lookback, chunk, lookahead] in 60ms frames
result = model.generate(
input="audio.wav",
chunk_size=chunk_size,
encoder_chunk_look_back=4,
decoder_chunk_look_back=1,
)
print(result[0]["text"])
| Property | Value |
|---|---|
| Architecture | Paraformer-Streaming |
| Parameters | 220M |
| Languages | Chinese, English |
| Mode | Streaming (online) |
| Sample Rate | 16kHz |
This repository publishes the model weights and accompanying files under the Apache License 2.0, unless an individual file carries a different notice. At the immutable apache-2.0-20260804 tag, this statement covers model.pt, config.yaml, configuration.json, am.mvn, seg_dict, tokens.json, the example audio, the architecture figure, and this model card.
For an auditable deployment, pin both the repository and revision:
model = AutoModel(
model="funasr/paraformer-zh-streaming",
model_revision="apache-2.0-20260804",
hub="hf",
device="cuda",
)
Record the resolved commit SHA and artifact checksums in deployment metadata. The historical revision 45df2405b92dfeb21e01ea4465954677005ae370 did not contain the license file referenced by its model card and remains unchanged; new integrations should use the immutable tag above. This clarification applies to the tagged snapshot and does not rewrite an older revision retroactively.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.
How it works