Claude
f9e1eb82c4
app/ = /opt/callrec (Node-Backend, Vite-Frontend, Python-Inferenz, Audio-Skripte, rkllm_server.cpp), system/ = systemd-Units, WirePlumber/PipeWire-Drop-ins, sudoers. Modelle, Binaries, Aufnahmen und node_modules bleiben draussen (.gitignore). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
67 lines
2.4 KiB
Python
67 lines
2.4 KiB
Python
#!/opt/aime-venv/bin/python
|
|
"""CallRec NPU transcriber: wav -> JSON words on stdout.
|
|
large-v3-turbo decoder on CPU (int8), encoder on RK3588 NPU (rknn fp16).
|
|
Output matches what recorder.js consumes: {"text","words":[{word,t_start_ms,t_end_ms,probability}]}.
|
|
Usage: transcribe_cli.py <wav> [--cpu] (--cpu = no NPU, plain CPU encoder for A/B)
|
|
"""
|
|
import sys, json, time
|
|
import numpy as np
|
|
|
|
WAV = sys.argv[1]
|
|
USE_NPU = "--cpu" not in sys.argv[2:]
|
|
import os
|
|
BATCH = int(os.environ.get("WHISPER_BATCH", "8"))
|
|
MODEL = "large-v3-turbo"
|
|
ROOT = "/opt/callrec/models/whisper"
|
|
RKNN = "/opt/callrec/models/turbo-encoder-fp16.rknn"
|
|
sys.path.insert(0, "/opt/callrec")
|
|
|
|
from faster_whisper import WhisperModel, BatchedInferencePipeline
|
|
|
|
t0 = time.time()
|
|
model = WhisperModel(MODEL, device="cpu", compute_type="int8",
|
|
download_root=ROOT, cpu_threads=4)
|
|
|
|
if USE_NPU:
|
|
from faster_whisper.transcribe import get_ctranslate2_storage
|
|
from inference.rknpu_whisper_encoder import RknnWhisperEncoder
|
|
npu = RknnWhisperEncoder(RKNN, lib_path="/usr/lib/librknnrt.so")
|
|
def npu_encode(features):
|
|
feats = np.asarray(features, dtype=np.float32)
|
|
out = npu.run(feats)
|
|
return get_ctranslate2_storage(np.ascontiguousarray(out))
|
|
model.encode = npu_encode
|
|
|
|
engine = BatchedInferencePipeline(model=model) if BATCH > 1 else model
|
|
t_load = time.time() - t0
|
|
t0 = time.time()
|
|
if BATCH > 1:
|
|
segments, info = engine.transcribe(WAV, language="de", beam_size=1,
|
|
batch_size=BATCH, word_timestamps=True)
|
|
else:
|
|
segments, info = engine.transcribe(WAV, language="de", beam_size=1,
|
|
word_timestamps=True, vad_filter=True)
|
|
words = []
|
|
texts = []
|
|
for seg in segments:
|
|
texts.append(seg.text.strip())
|
|
for w in (seg.words or []):
|
|
words.append({
|
|
"word": w.word.strip(),
|
|
"t_start_ms": int(round((w.start or 0) * 1000)),
|
|
"t_end_ms": int(round((w.end or 0) * 1000)),
|
|
"probability": round(w.probability, 3),
|
|
})
|
|
t_trans = time.time() - t0
|
|
out = {
|
|
"text": " ".join(texts).strip(),
|
|
"words": words,
|
|
"duration": round(float(info.duration or 0), 1),
|
|
"load_s": round(t_load, 1),
|
|
"transcribe_s": round(t_trans, 1),
|
|
"rtf": round(t_trans / max(float(info.duration or 0.1), 0.1), 3),
|
|
"compute": "npu" if USE_NPU else "cpu",
|
|
"batch": BATCH,
|
|
}
|
|
print(json.dumps(out, ensure_ascii=False))
|