bosonai_higgs-audio-v3-8b-stt-v2

---

license: apache-2.0 language:

• en tags:
• automatic-speech-recognition
• hf-asr-leaderboard
• whisper
• qwen pipeline_tag: automatic-speech-recognition base_model: bosonai/higgs-audio-v3-8b model-index:
• name: higgs-audio-v3-8b-stt-v2 results:
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: AMI (Meetings test) type: edinburghcstr/ami config: ihm split: test args: language: en metrics:
    • name: Test WER type: wer value: 10.14
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: Earnings-22 type: revdotcom/earnings22 split: test args: language: en metrics:
    • name: Test WER type: wer value: 8.73
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: GigaSpeech type: speechcolab/gigaspeech split: test args: language: en metrics:
    • name: Test WER type: wer value: 8.47
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: LibriSpeech (clean) type: librispeech_asr config: clean split: test args: language: en metrics:
    • name: Test WER type: wer value: 1.25
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: LibriSpeech (other) type: librispeech_asr config: other split: test args: language: en metrics:
    • name: Test WER type: wer value: 2.38
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: SPGI Speech type: kensho/spgispeech config: test split: test args: language: en metrics:
    • name: Test WER type: wer value: 3.60
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: tedlium-v3 type: LIUM/tedlium config: release1 split: test args: language: en metrics:
    • name: Test WER type: wer value: 3.09
  • task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: Vox Populi type: facebook/voxpopuli config: en split: test args: language: en metrics:
    • name: Test WER type: wer value: 5.92

---

Higgs Audio v3 8B STT v2

A speech-to-text model combining a Whisper-Large-v3 encoder with a Qwen3-8B decoder (8.91B total parameters), fine-tuned with LoRA on diverse ASR benchmarks.

Usage

import torch
import numpy as np
from transformers import AutoModel, AutoTokenizer

# Load model
model = AutoModel.from_pretrained(
    "bosonai/higgs-audio-v3-8b-stt-v2",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    attn_implementation="eager",
    device_map="cuda:0",
)
tokenizer = AutoTokenizer.from_pretrained("bosonai/higgs-audio-v3-8b-stt-v2")

# Transcribe audio (16kHz mono numpy array)
from transformers.utils import cached_file
import importlib.util
spec = importlib.util.spec_from_file_location("transcribe", cached_file("bosonai/higgs-audio-v3-8b-stt-v2", "transcribe.py", _raise_exceptions_for_connection_errors=False))
mod = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)

audio_np = np.random.randn(16000).astype(np.float32)  # replace with your audio
text = mod.transcribe(model, tokenizer, audio_np)
print(text)

Requirements

torch
transformers>=4.51.0
whisper  # for audio preprocessing (WhisperProcessor)

Architecture

• Encoder: Whisper-Large-v3 (frozen)
• Decoder: Qwen3-8B (LoRA fine-tuned, merged)
• Total parameters: 8.91B
• Audio input: 16kHz mono WAV
• Supports: Thinking mode for improved accuracy

Performance (ESB Benchmark — Full Scale, All Samples)

Dataset	WER
AMI	10.14%
Earnings22	8.73%
GigaSpeech	8.47%
LibriSpeech Clean	1.25%
LibriSpeech Other	2.38%
SPGISpeech	3.60%
TED-LIUM	3.09%
VoxPopuli	5.92%
Average	5.449%