Automatic Speech Recognition
PEFT
Safetensors
English
whisper
lora
audio
machine-sound
anomaly-detection
predictive-maintenance
manufacturing
voice-agent
text-to-speech
non-commercial
Eval Results (legacy)
Instructions to use muruganmath/Machine-fault-detector with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use muruganmath/Machine-fault-detector with PEFT:
from peft import PeftModel from transformers import AutoModelForSpeechSeq2Seq base_model = AutoModelForSpeechSeq2Seq.from_pretrained("openai/whisper-large-v3") model = PeftModel.from_pretrained(base_model, "muruganmath/Machine-fault-detector") - Notebooks
- Google Colab
- Kaggle
Add trained LoRA adapter (main), tokenizer/processor, lineage, standalone inference and model card
36ba721 verified Download diagnosis.py from muruganmath/Machine-fault-detector: direct link, hf CLI and curl.
- Browser
- Download file 4.44 kB
-
https://huggingface.co/muruganmath/Machine-fault-detector/resolve/main/diagnosis.py
- Command line
-
hf download hf://muruganmath/Machine-fault-detector/diagnosis.py
-
curl -L -o diagnosis.py https://huggingface.co/muruganmath/Machine-fault-detector/resolve/main/diagnosis.py
4.44 kB
| """Machine-sound diagnosis from the whisper_ft adapter -> a sentence the agent can speak. | |
| The whisper_ft main adapter (whisper_ft/out/whisper-lora/main) was trained to describe | |
| non-speech audio with one of three templates (whisper_ft/configs/templates.yaml): | |
| anomaly: "machine: fan; condition: high load; status: anomalous; fault: overheating; noise: clean" | |
| engine: "engine sound; rpm mean 2020 range 920 to 3860; torque mean 467 nm" | |
| ego: "activity: ...; task: ...; view: ...[; notes: ...]" | |
| parse_diagnosis() turns the decoded text into fields (None if it is not a complete template), | |
| describe() turns the fields into plain English for TTS, e.g. | |
| "Sound check. This is a fan under high load. It sounds abnormal. Possible fault: overheating." | |
| Wording was chosen by synthesizing with MMS-TTS and transcribing back with Whisper (2026-11-12): | |
| MMS cannot say "engine" ("Insan", "Ed Sheik") or "idling" ("piling"); "machine sound check" | |
| came back as "matching sound check". "motor", "at idle" and "Sound check. This is ..." survive. | |
| The diagnosis is the model's estimate (anomaly F1 1.0, field accuracy 0.60 on validation, | |
| whisper_ft train_summary.json): describe() says "possible" fault, never a certainty. | |
| """ | |
| from __future__ import annotations | |
| import re | |
| from dataclasses import dataclass, field | |
| class Diagnosis: | |
| kind: str # "anomaly" | "engine" | "activity" | |
| fields: dict = field(default_factory=dict) | |
| raw: str = "" | |
| _ENGINE = re.compile( | |
| r"^\s*engine sound\s*;\s*rpm mean\s+(-?\d+)\s+range\s+(-?\d+)\s+to\s+(-?\d+)\s*(?:;(.*))?$", | |
| re.IGNORECASE | re.DOTALL) | |
| _TORQUE = re.compile(r"^\s*torque mean\s+(-?\d+)\s*nm\s*\.?\s*$", re.IGNORECASE) | |
| def _kv(text: str) -> dict: | |
| out = {} | |
| for part in text.split(";"): | |
| if ":" in part: | |
| k, v = part.split(":", 1) | |
| out[k.strip().lower()] = v.strip().strip(".").lower() | |
| return out | |
| def parse_diagnosis(text: str) -> Diagnosis | None: | |
| text = (text or "").strip() | |
| m = _ENGINE.match(text) | |
| if m: | |
| mean, lo, hi = (int(g) for g in m.groups()[:3]) | |
| f = {"rpm_mean": mean, "rpm_min": lo, "rpm_max": hi} | |
| t = _TORQUE.match(m.group(4) or "") | |
| if t: # a garbled torque ("6-7-0-8 .", seen 2026-11-12) is dropped, not guessed | |
| f["torque_mean"] = int(t.group(1)) | |
| if not lo <= mean <= hi: # inconsistent numbers: not a usable diagnosis | |
| return None | |
| return Diagnosis("engine", f, text) | |
| f = _kv(text) | |
| if f.get("machine") and f.get("condition") and f.get("status") in {"normal", "anomalous"}: | |
| return Diagnosis("anomaly", f, text) | |
| if f.get("activity") and f.get("task"): | |
| return Diagnosis("activity", f, text) | |
| return None | |
| def _article(noun: str) -> str: | |
| return ("an " if noun[:1] in "aeiou" else "a ") + noun | |
| def _condition(c: str) -> str: | |
| if c == "idle": | |
| return "at idle" | |
| if c.endswith("load"): | |
| return "under " + c | |
| return "in " + c + " condition" | |
| def _num(n: int) -> str: | |
| return ("minus " + str(-n)) if n < 0 else str(n) | |
| def describe(d: Diagnosis) -> str: | |
| f = d.fields | |
| if d.kind == "anomaly": | |
| s = f"Sound check. This is {_article(f['machine'])} {_condition(f['condition'])}." | |
| if f["status"] == "anomalous": | |
| s += " It sounds abnormal." | |
| fault = f.get("fault", "none") | |
| if fault and fault != "none": | |
| s += f" Possible fault: {fault}." | |
| else: | |
| s += " It sounds normal." | |
| noise = f.get("noise", "") | |
| if noise and noise not in {"clean", "none"}: | |
| s += f" The recording has {noise.replace(' noise', '')} background noise." | |
| return s | |
| if d.kind == "engine": | |
| # "estimated": the adapter's numbers can be far off (505_Engine-C: said mean 4660 RPM, label 2870) | |
| s = (f"Sound check. This is a motor. The estimated average speed is {_num(f['rpm_mean'])} RPM, " | |
| f"ranging from {_num(f['rpm_min'])} to {_num(f['rpm_max'])} RPM.") | |
| if "torque_mean" in f: | |
| s += f" The estimated average torque is {_num(f['torque_mean'])} newton metres." | |
| return s | |
| s = f"Sound check. Workshop activity: {f['activity']}. Task: {f['task']}." | |
| if f.get("view"): | |
| s += f" View: {f['view']}." | |
| if f.get("notes"): | |
| s += f" Notes: {f['notes']}." | |
| return s | |