mythicinfinity/speech-truncation-detection-12M

Speech Truncation Detection (12M)

Frame-wise speech truncation detector for identifying whether audio ends in a likely cutoff.

The model evaluates the tail of each input waveform and produces per-frame truncation probabilities. A final clip-level score is computed from the final portion of the sequence, then thresholded to return a boolean is_truncated prediction.

Intended Use

This model is intended for speech recordings where you want to decide whether the spoken audio appears to end abruptly (truncated) versus ending naturally.

Use this for speech-oriented pipelines such as dataset quality checks, ingestion validation, or gating before downstream ASR/transcript correction steps.

This model is not intended for non-speech audio classification tasks (music, environmental sounds, general acoustic event detection).

Quick Start

import numpy as np
from transformers import AutoModel

repo_id = "mythicinfinity/speech-truncation-detection-12M"
model = AutoModel.from_pretrained(repo_id, trust_remote_code=True)
model.eval()

sr = 16000
wav = np.zeros(sr * 2, dtype=np.float32)
out = model.predict_from_audio(audio=wav, sampling_rate=sr)
is_truncated = bool(out.is_truncated[0].item())
confidence = float(out.confidence_score[0].item())
print(is_truncated, confidence)

Usage Examples

Single waveform

import numpy as np
from transformers import AutoModel

model = AutoModel.from_pretrained("mythicinfinity/speech-truncation-detection-12M", trust_remote_code=True)
waveform = np.random.randn(32000).astype('float32')  # ~2s at 16kHz
out = model.predict_from_audio(audio=waveform, sampling_rate=16000)
print(float(out.confidence_score[0]), bool(out.is_truncated[0]))

Batched tensor of waveforms ([batch, time])

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("mythicinfinity/speech-truncation-detection-12M", trust_remote_code=True)
audio_batch = torch.randn(4, 32000)  # 4 clips, each ~2s at 16kHz
out = model.predict_from_audio(audio=audio_batch, sampling_rate=16000)
print(out.confidence_score.shape)  # torch.Size([4])
print(out.is_truncated.shape)      # torch.Size([4])

List of waveforms (variable lengths) with per-item sample rates

import numpy as np
from transformers import AutoModel

model = AutoModel.from_pretrained("mythicinfinity/speech-truncation-detection-12M", trust_remote_code=True)
audio_list = [
    np.random.randn(24000).astype('float32'),  # 1.5s @ 16kHz
    np.random.randn(8000).astype('float32'),   # 1.0s @ 8kHz
]
sample_rates = [16000, 8000]
out = model.predict_from_audio(audio=audio_list, sampling_rate=sample_rates)
print(out.confidence_score)
print(out.is_truncated)

Accepted input forms: single tensor/ndarray waveform, batched tensor ([B, T]), or list/tuple of waveforms.

Outputs

• truncation_probabilities: per-frame probability of truncation ([batch, time])
• confidence_score: clip-level truncation score ([batch])
• is_truncated: clip-level boolean decision ([batch])

Non-Remote-Code Fallback

If you prefer not to use trust_remote_code=True, download the artifact files (config.json, model.safetensors, configuration_*.py, modeling_*.py, processing_*.py) into a local directory and load from those local files:

import json
import sys
from pathlib import Path
from safetensors.torch import load_file

artifact_dir = Path('/path/to/downloaded-artifact')
sys.path.insert(0, str(artifact_dir))

from configuration_speech_truncation_detection import SpeechTruncationDetectionConfig
from modeling_speech_truncation_detection import SpeechTruncationDetectionModel

cfg_payload = json.loads((artifact_dir / 'config.json').read_text(encoding='utf-8'))
cfg = SpeechTruncationDetectionConfig(**cfg_payload)
model = SpeechTruncationDetectionModel(cfg)
state = load_file(str(artifact_dir / 'model.safetensors'), device='cpu')
model.load_state_dict(state, strict=True)
model.eval()