fasr-asr-paraformer 0.5.2

paraformer asr model for fasr

fasr-asr-paraformer

Chinese documentation

Paraformer speech recognition for fasr. The offline models return timestamped tokens, and seaco_paraformer adds hotword support.

Install

pip install fasr-asr-paraformer

Registered Models

Registry name	Class	Best for
paraformer	Paraformer	Offline ASR with token timestamps
seaco_paraformer	SeacoParaformer	Offline ASR with hotword biasing
paraformer_online	ParaformerOnline	Streaming Paraformer ASR

Pipeline Usage

from fasr import AudioPipeline

pipeline = (
    AudioPipeline()
    .add_pipe("detector", model="fsmn")
    .add_pipe(
        "recognizer",
        model="paraformer",
        batch_size=64,
        disable_log=True,
    )
    .add_pipe("sentencizer", model="ct_transformer")
)

For hotwords, use seaco_paraformer:

pipeline = (
    AudioPipeline()
    .add_pipe("detector", model="fsmn")
    .add_pipe("recognizer", model="seaco_paraformer")
    .add_pipe("sentencizer", model="ct_transformer")
)

audio = pipeline.run("meeting.wav", hotwords=["Paraformer", "fasr"])[0]

Confection Config

[asr_model]
@asr_models = "paraformer"
batch_size = 64
device = "cuda:0"
disable_update = true
disable_log = true
disable_pbar = true

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 = "paraformer"
batch_size = 64
device = "cuda:0"
disable_log = true
disable_pbar = true

Direct Model Usage

from fasr.config import registry

model = registry.asr_models.get("paraformer")(batch_size=64)
spans = model.transcribe(audio_spans)
for span in spans:
    print(span.text)
    for token in span.tokens or []:
        print(token.text, token.start_ms, token.end_ms)

Use local weights:

model.load_checkpoint("/path/to/paraformer")

Parameters

Parameter	Type / range	Default	Higher / true	Lower / false	Change when
batch_size	int >= 1	10000	More throughput, more memory	Lower memory, lower throughput	Batch size is too memory-heavy or too slow
device	str | null	null	Explicit device selection such as cuda:0 or cpu	Auto-selects cuda:0 when CUDA is available, otherwise cpu	You want to override the automatic device choice
disable_update	bool	True	Skips FunASR update checks	Allows update checks	Reproducible startup or update discovery
disable_log	bool	True	Suppresses backend logs	Shows backend logs	Debugging loading or inference
disable_pbar	bool	True	Hides progress bars	Shows progress bars	Interactive scripts
compile_model	bool, online only	False	Uses torch.compile, slower warm-up but faster steady state	No compile warm-up	Long-running streaming service
chunk_size_ms	int, online only	600	Fewer streaming calls, later output	More responsive, more overhead	Streaming latency/throughput tuning

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

Output

• paraformer and seaco_paraformer populate span.raw_text.
• When timestamps are available, span.tokens contains absolute start_ms / end_ms values on the channel timeline.
• If device is omitted, the model auto-selects cuda:0 when CUDA is available, otherwise it falls back to cpu and logs a warning.
• seaco_paraformer accepts hotwords through the pipeline run call or transcribe(..., hotwords=[...]).

Dependencies

• fasr
• funasr
• Python 3.10-3.12