简体中文 | English
Choose the model, checkpoint format, and runtime separately. A model listed here is not automatically supported by every server or export backend.
| Task | Model family | Important boundary |
|---|---|---|
| Context-aware file transcription | Fun-ASR-Nano | The base checkpoint and native vLLM-converted checkpoint are different artifacts. |
| Broader multilingual file transcription | Fun-ASR-MLT-Nano | Separate checkpoint; do not attribute its language coverage to the base Nano model. |
| Transcription with emotion and audio-event tags | SenseVoiceSmall | Tags are model outputs, not speaker identity. Speaker-aware pipelines need the documented companion components. |
| Mandarin transcription with timestamps | Paraformer | Offline and streaming checkpoints have different inference contracts. |
| Offline text, timestamps and anonymous speakers together | MOSS-Transcribe-Diarize | Third-party OpenMOSS model; no external VAD or speaker model is needed for its unified path. Not known-person identification. |
Use the selection guide for workload choices, the SDK contract for parameters and return values, and the deployment matrix for serving options.
Use FunAudioLLM/Fun-ASR-Nano-2512-hf
with AutoProcessor and AutoModelForSpeechSeq2Seq, not the original toolkit,
native vLLM or GGUF loader. Installation and inference
use released Transformers 5.17.0 and a pinned official revision. CPU examples,
batching and a notebook are available; matching torchaudio is required. This
zh/en/ja export returns text, not word timestamps or speaker identities.
Start with the installation guide and Python tutorial. Use an explicit hub and record the resolved checkpoint/revision, FunASR version, device and inference options. Aliases are resolved by the repository mapping; an alias is convenient, but is not an immutable model revision.
from funasr import AutoModel
model = AutoModel(model="paraformer-zh", hub="ms", device="cpu")
result = model.generate(input="meeting.wav")
print(result[0]["text"])Replace meeting.wav with an existing recording. Download time, warmup and
inference are separate measurements. Preserve the raw result when validating
timestamps, speaker labels or model-specific tags.
| SDK alias | Intended use | ModelScope (hub="ms") |
Hugging Face (hub="hf") |
|---|---|---|---|
paraformer-zh |
Offline Mandarin transcription; ModelScope alias selects SeACo | SeACo checkpoint | Checkpoint |
paraformer-zh-streaming |
Chunked streaming with per-session cache | Checkpoint | Checkpoint |
paraformer-en |
Offline English transcription | Resolve through the hub mapping. | Resolve through the hub mapping. |
The legacy Paraformer VAD/punctuation pipeline
and paraformer-zh-spk listing describe pipelines, not the paraformer-zh alias target. For explicit
composition, use vad_model, punc_model and spk_model as documented in the
SDK guide, rather than treating all components as one
interchangeable ASR checkpoint.
For speaker vectors and SenseVoice raw/display tags, use the speaker and emotion guide. For utterance-end KWS, use the separate keyword spotting guide.
| Component | Alias | Model cards | What it does not do |
|---|---|---|---|
| Voice activity detection | fsmn-vad |
ModelScope / HF | Does not transcribe speech or identify a speaker. |
| Punctuation restoration | ct-punc |
ModelScope / HF | Does not create acoustic timestamps. |
| Speaker embeddings | cam++ |
Hub mapping | Does not name known people without a separately designed enrollment/matching system. |
| Timestamp prediction | fa-zh |
Hub mapping | Must match the documented input/model path; not universal timestamp support for every recognizer. |
See the full ModelScope inventory and Hugging Face inventory for additional checkpoints. These inventories include historical variants; verify each model card before using a checkpoint in a new service.
MOSS-Transcribe-Diarize is published by OpenMOSS, not by the FunASR team. Its unified offline output includes text, timestamps and anonymous speaker labels scoped to a recording. It is not a realtime streaming or known-person identification model. Use the MOSS guide for the adapter, native upstream servers, memory requirements and response boundaries.
The FunASR software license does not grant one license for every model weight. Consult the individual checkpoint's model card/license, publisher, training-data notes and the applicable Model License Agreement. Keep upstream attribution when redistributing models or derived artifacts.
- Train or fine-tune a supported recipe.
- Register a custom model.
- Choose native vLLM or split-engine; preserve their distinct artifacts and API contracts.
- Deploy GGUF models with llama.cpp; an ONNX export is not a GGUF conversion.
- Measure quality and runtime separately, using representative audio and exact versions.