istrain
559caead36
Public community release. The complete working system — DSP + decoders (scripts/), web dashboard + API (dashboard/), container stack (docker/), config templates (config/) — plus APOCALYPSE-EDITION.md, the full build guide for a human or a coding agent. GPLv3 (PyEOT dependency). Scrubbed of all secrets, internal IPs, hostnames, and location detail. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
98 lines
4.5 KiB
Python
98 lines
4.5 KiB
Python
#!/usr/bin/env python3
|
|
"""istrain — voice transcription worker (VM side).
|
|
|
|
Watches the airband capture dir for new voice clips and sends each to the Mill's
|
|
whisper service (istrain-whisper stack, Dockge on the Mill). CPU lives on the Mill;
|
|
this worker only enhances (ffmpeg comms filter) and ships. Results land as:
|
|
- a sidecar <clip>.txt next to the mp3 (same convention as radio/transcribe)
|
|
- a durable line in ~/istrain/transcripts.jsonl (clips in /tmp are ephemeral;
|
|
the transcript record survives reboots — keep-everything policy)
|
|
|
|
Lineage: projects/radio/transcribe/transcribe-worker.sh (marine post) — whisper.cpp
|
|
run locally there; here the model runs on the Mill and VAD replaces the prompt-echo
|
|
fixes (radio/transcribe/NOTES.md — VAD was the named "robust upgrade").
|
|
"""
|
|
import json, os, subprocess, sys, time, urllib.request, urllib.error, uuid
|
|
|
|
WATCH = os.environ.get("CLIPS_DIR", "/tmp/airband")
|
|
MILL = os.environ.get("MILL_ASR", "http://whisper:9000")
|
|
ASR = MILL + "/asr?task=transcribe&language=en&vad_filter=true&output=txt"
|
|
# env-overridable for the Mill container (writes /data/transcripts.jsonl); VM default unchanged.
|
|
OUT_LOG = os.path.expanduser(os.environ.get("TRANSCRIPTS_LOG", "~/istrain/transcripts.jsonl"))
|
|
MIN_BYTES = 5000 # below this it's a squelch blip (marine rule)
|
|
SETTLE_S = 5 # still being written if newer than this
|
|
POLL_S = 10
|
|
NO_SPEECH = "(no speech)"
|
|
|
|
def enhance(src: str, dst: str) -> bool:
|
|
"""Comms-audio cleanup: bandpass to voice, normalize. 16k mono wav."""
|
|
r = subprocess.run(
|
|
["ffmpeg", "-hide_banner", "-loglevel", "error", "-y", "-i", src,
|
|
"-af", "highpass=f=250,lowpass=f=3200,speechnorm=e=12:r=0.0001:l=1",
|
|
"-ar", "16000", "-ac", "1", dst],
|
|
capture_output=True)
|
|
return r.returncode == 0 and os.path.exists(dst)
|
|
|
|
def transcribe(wav: str) -> str | None:
|
|
"""POST the clip to the Mill. Returns text ('' if no speech), None on error."""
|
|
boundary = uuid.uuid4().hex
|
|
with open(wav, "rb") as f:
|
|
payload = f.read()
|
|
body = (f"--{boundary}\r\nContent-Disposition: form-data; "
|
|
f'name="audio_file"; filename="{os.path.basename(wav)}"\r\n'
|
|
f"Content-Type: audio/wav\r\n\r\n").encode() + payload + f"\r\n--{boundary}--\r\n".encode()
|
|
req = urllib.request.Request(ASR, data=body, method="POST",
|
|
headers={"Content-Type": f"multipart/form-data; boundary={boundary}"})
|
|
try:
|
|
with urllib.request.urlopen(req, timeout=240) as resp:
|
|
text = resp.read().decode().strip()
|
|
except (urllib.error.URLError, TimeoutError) as e:
|
|
print(f"ASR error: {e}", flush=True)
|
|
return None
|
|
# whisper renders non-speech as runs of "..." — normalize to empty
|
|
cleaned = " ".join(t for t in text.split() if set(t) - set(".")).strip()
|
|
return cleaned
|
|
|
|
def main():
|
|
print(f"transcribe-worker: watching {WATCH}, ASR={MILL}", flush=True)
|
|
while True:
|
|
try:
|
|
# newest first: fresh clips (what the live panel + push want) never wait behind backlog
|
|
names = sorted(os.listdir(WATCH),
|
|
key=lambda n: os.path.getmtime(os.path.join(WATCH, n))
|
|
if os.path.exists(os.path.join(WATCH, n)) else 0, reverse=True)
|
|
except FileNotFoundError:
|
|
time.sleep(POLL_S); continue
|
|
now = time.time()
|
|
for name in names:
|
|
if not name.endswith(".mp3"):
|
|
continue
|
|
mp3 = os.path.join(WATCH, name)
|
|
txt = mp3[:-4] + ".txt"
|
|
if os.path.exists(txt):
|
|
continue
|
|
try:
|
|
st = os.stat(mp3)
|
|
except FileNotFoundError:
|
|
continue
|
|
if st.st_size < MIN_BYTES or now - st.st_mtime < SETTLE_S:
|
|
continue
|
|
wav = f"/tmp/transcribe-{os.getpid()}.wav"
|
|
ok = enhance(mp3, wav)
|
|
text = transcribe(wav if ok else mp3)
|
|
if os.path.exists(wav):
|
|
os.unlink(wav)
|
|
if text is None: # service unreachable — retry next poll, don't mark
|
|
break
|
|
with open(txt, "w") as f:
|
|
f.write(text if text else NO_SPEECH)
|
|
if text: # only real speech goes in the durable log
|
|
rec = {"ts": st.st_mtime, "file": name, "size": st.st_size, "text": text}
|
|
with open(OUT_LOG, "a") as f:
|
|
f.write(json.dumps(rec) + "\n")
|
|
print(f"TRANSCRIBED {name}: {text[:100]}", flush=True)
|
|
time.sleep(POLL_S)
|
|
|
|
if __name__ == "__main__":
|
|
main()
|