fasr-asr-firered 0.5.2

FireRed ASR for fasr (bundled fireredasr2 inference)

fasr-asr-firered

Chinese documentation

FireRedASR2 speech recognition for fasr. The plugin exposes both AED decoding and LLM decoding. AED can return token timestamps; LLM focuses on full-text accuracy without timestamps.

Install

pip install fasr-asr-firered

Registered Models

Registry name	Class	Best for
firered	FireRedAEDForASR	Default alias for AED mode
firered_aed	FireRedAEDForASR	Timestamped AED recognition
firered_llm	FireRedLLMForASR	LLM decoding, no timestamps

Default checkpoints:

Model	Checkpoint
firered_aed	FireRedTeam/FireRedASR2-AED
firered_llm	FireRedTeam/FireRedASR2-LLM

Pipeline Usage

from fasr import AudioPipeline

pipeline = (
    AudioPipeline()
    .add_pipe("detector", model="fsmn")
    .add_pipe(
        "recognizer",
        model="firered_aed",
        device="cuda",
        beam_size=3,
        return_timestamp=True,
    )
    .add_pipe("sentencizer", model="ct_transformer")
)

Quick choices:

Goal	Use	Result
Token timestamps	model="firered_aed", return_timestamp=True	Populates span.tokens
Full-text decoding	model="firered_llm"	Populates span.raw_text, no timestamps
Lower VRAM for AED	use_half=True	FP16 inference on GPU
CPU inference	device="cpu"	Runs without CUDA, slower
Wider search	beam_size=5	Potentially better accuracy, slower

Confection Config

[asr_model]
@asr_models = "firered_aed"
device = "cuda"
beam_size = 3
return_timestamp = true
use_half = true

Inside a pipeline:

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

[pipeline.pipes]

[pipeline.pipes.recognizer]
@pipes = "thread_pipe"
batch_size = 2

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

[pipeline.pipes.recognizer.component.model]
@asr_models = "firered_aed"
device = "cuda"
beam_size = 3
return_timestamp = true

Direct Model Usage

from fasr.config import registry

model = registry.asr_models.get("firered_aed")(
    device="cuda",
    beam_size=3,
    return_timestamp=True,
)

spans = model.transcribe(audio_spans)
for span in spans:
    print(span.text)

Use local weights:

model.load_checkpoint("/path/to/FireRedASR2-AED")

Shared Parameters

Parameter	Type / range	Default	Higher / true	Lower / false	Change when
device	str or None	None	"cuda" uses GPU	"cpu" uses CPU	Deployment target changes
beam_size	int >= 1	3	Wider search, slower, more memory	Faster, possibly lower accuracy	Accuracy/speed tradeoff
decode_max_len	int >= 0	0	Allows longer outputs	Shorter cap; 0 lets backend decide	Output is truncated or too long

AED Parameters

Parameter	Type / range	Default	Higher / true	Lower / false	Change when
use_half	bool	True	Lower VRAM, faster on GPU	FP32, more stable	GPU memory or numeric stability matters
nbest	int >= 1	1	More hypotheses	Single best result	You need alternative hypotheses
softmax_smoothing	float	1.25	Smoother distribution	Sharper distribution	Beam search needs tuning
aed_length_penalty	float	0.6	Favors different output lengths	Less length adjustment	Output is too short or too long
eos_penalty	float	1.0	Discourages ending too early	Easier EOS	Decoding ends too early or too late
return_timestamp	bool	True	Returns token timestamps	Text only	You need word/character timing
elm_weight	float	0.0	More external LM influence	0.0 disables external LM	You provide elm_dir

LLM Parameters

Parameter	Type / range	Default	Higher value	Lower value	Change when
decode_min_len	int >= 0	0	Forces longer minimum output	Allows shorter output	Output ends too early
repetition_penalty	float	1.2	Stronger repetition suppression	Allows more repetition	Repeated phrases appear
llm_length_penalty	float	0.0	Adjusts length preference	Less length adjustment	Output length is biased
temperature	float >= 0	1.0	More diverse, less deterministic	More deterministic	You need stability or diversity

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

Output

• AED writes span.raw_text.
• AED also fills span.tokens when return_timestamp=True.
• LLM writes span.raw_text and leaves span.tokens empty.

Dependencies

• fasr
• torch >= 2.0.0
• torchaudio
• transformers >= 4.36
• librosa >= 0.10.0
• kaldiio >= 2.18.0
• kaldi-native-fbank >= 1.19.0
• Python 3.10-3.12