fasr-asr-funasr 0.5.2

Fun ASR model for fasr

fasr-asr-funasr

Chinese documentation

Fun-ASR-Nano speech recognition for fasr. This plugin wraps funasr.AutoModel and writes whole-utterance text to AudioSpan.raw_text. It does not produce word or character timestamps.

Install

pip install fasr-asr-funasr

Registered Model

Registry name	Class	Best for
fun_asr_nano	FunASRNano	Lightweight FunASR recognition without timestamps

The default checkpoint is FunAudioLLM/Fun-ASR-Nano-2512.

Pipeline Usage

from fasr import AudioPipeline

pipeline = (
    AudioPipeline()
    .add_pipe("detector", model="fsmn")
    .add_pipe(
        "recognizer",
        model="fun_asr_nano",
        device="cuda:0",
        language="中文",
        itn=True,
    )
    .add_pipe("sentencizer", model="ct_transformer")
)

Quick choices:

Goal	Use	Result
CPU inference	device="cpu"	Runs without CUDA, usually slower
First GPU	device="cuda:0"	Uses GPU 0
Disable number/date normalization	itn=False	Keeps raw spoken text form
Force language hint	language="中文"	Passes the language label to FunASR

Confection Config

[asr_model]
@asr_models = "fun_asr_nano"
device = "cuda:0"
language = "中文"
itn = true
batch_size = 1

Inside a pipeline:

[pipeline]
@pipelines = "AudioPipeline.v1"
pipe_order = ["recognizer"]

[pipeline.pipes]

[pipeline.pipes.recognizer]
@pipes = "thread_pipe"
batch_size = 4

[pipeline.pipes.recognizer.component]
@components = "recognizer"

[pipeline.pipes.recognizer.component.model]
@asr_models = "fun_asr_nano"
device = "cuda:0"
language = "中文"
itn = true
batch_size = 1

Direct Model Usage

from fasr.config import registry

model = registry.asr_models.get("fun_asr_nano")()
model.device = "cuda:0"

# `transcribe` mutates and returns the input AudioSpan list.
spans = model.transcribe(audio_spans, hotwords=["fasr"], language="中文")
for span in spans:
    print(span.text)

Use local weights:

model.load_checkpoint("/path/to/Fun-ASR-Nano")

Parameters

Parameter	Type / range	Default	Higher / true	Lower / false	Change when
device	str	"cuda:0"	Selects a CUDA device such as "cuda:1"	"cpu" runs on CPU	Deployment target changes
language	str	"中文"	Stronger language hint	Use another supported label	Audio language is known
itn	bool	True	Normalizes numbers, dates, and similar text	Keeps less-normalized text	You need spoken-form output
batch_size	int >= 1	1	Better throughput, more memory	Lower memory, less throughput	Batch inference needs tuning
trust_remote_code	bool	True	Allows custom checkpoint code	Safer, but some checkpoints may fail	Running untrusted checkpoints

Generic checkpoint fields such as checkpoint, cache_dir, endpoint, revision, and force_download are inherited from the base model.

Output

• span.raw_text is populated.
• span.tokens is not populated because Fun-ASR-Nano does not return timestamps.
• span.text reads the recognized text.

Dependencies

• fasr
• funasr
• Python 3.10-3.12