arc-tagger / stream_deepgram.py
ksingla025's picture
Arc Tagger v0.4.0
83a951c
Raw History Blame Contribute Delete
4.08 kB
"""Deepgram live ASR -> Arc Tagger, printed as JSON lines.
export DEEPGRAM_API_KEY=... # read from the environment only
pip install websockets # + sounddevice for --mic
python stream_deepgram.py --wav call.wav --domain talk
python stream_deepgram.py --mic --domain assistant --seconds 30
Interim results (is_final=false) are tagged immediately and revised when Deepgram revises them; is_final freezes the words;
speech_final / UtteranceEnd end the utterance (utterance-level intent, act, safety and role read-outs). With --diarize
(default on) Deepgram speaker ids are passed to the tagger, which enables speaker roles.
Each output line is {"type": "update", "source": "interim"|"final"|"endpoint", ...summary} (see arctag.asr.summarize);
the last line is {"type": "done", ...}. --result writes the full result (same schema as ArcTagger.tag()).
"""
import argparse, asyncio, json, os, sys, urllib.parse
HERE = os.path.dirname(os.path.abspath(__file__)); sys.path.insert(0, HERE)
from arctag import live # noqa: E402
from arctag.asr import DeepgramAdapter, summarize # noqa: E402
URL = "wss://api.deepgram.com/v1/listen"
def dg_url(a, sr):
q = dict(model=a.dg_model, language="en", encoding="linear16", sample_rate=sr, channels=1, interim_results="true",
punctuate="true", smart_format="true", diarize="true" if a.diarize else "false", vad_events="true",
utterance_end_ms=a.utterance_end_ms, endpointing=a.endpointing)
return URL + "?" + urllib.parse.urlencode(q)
async def connect(url, key):
import websockets
hdr = {"Authorization": "Token " + key}
try:
return await websockets.connect(url, additional_headers=hdr, max_size=None)
except TypeError: # websockets < 14
return await websockets.connect(url, extra_headers=hdr, max_size=None)
async def run(a):
key = live.api_key("DEEPGRAM_API_KEY")
from arctag import ArcTagger
tg = ArcTagger.from_pretrained(a.model_dir)
st = tg.stream(domain=a.domain, tier=a.tier)
ad = DeepgramAdapter(st)
sr, chunks = live.source(a)
try:
ws = await connect(dg_url(a, sr), key)
except Exception as e:
sys.exit("Deepgram connect failed: " + live.scrub(e, key))
async def send():
async for b in chunks:
await ws.send(b)
await ws.send(json.dumps({"type": "CloseStream"}))
async def recv():
async for raw in ws:
if isinstance(raw, bytes):
continue
m = json.loads(raw)
if m.get("type") == "Error":
live.emit({"type": "error", "message": live.scrub(m.get("description") or m.get("message") or "", key)})
continue
src = ("endpoint" if m.get("type") == "UtteranceEnd" else "final" if m.get("is_final") else "interim")
for up in ad(m):
if a.no_interim and src == "interim":
continue
live.emit(dict(type="update", source="endpoint" if up.get("utterances") else src, **summarize(up, st)))
tx = asyncio.create_task(send())
try:
await recv()
finally:
tx.cancel()
res = st.finalize()
live.emit({"type": "done", "n_words": len(res["words"]), "redacted_text": res["redacted_text"],
"latency_ms": st.s.stats()})
if a.result:
json.dump(res, open(a.result, "w"), indent=1, default=str)
def main():
ap = live.add_common_args(argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter))
ap.add_argument("--dg-model", default="nova-3")
ap.add_argument("--no-diarize", dest="diarize", action="store_false")
ap.add_argument("--endpointing", type=int, default=300, help="ms of silence for speech_final")
ap.add_argument("--utterance-end-ms", type=int, default=1000)
try:
asyncio.run(run(ap.parse_args()))
except KeyboardInterrupt:
pass
if __name__ == "__main__":
main()