"""Offline tests of the live ASR adapters against recorded-format provider messages (tests/fixtures, made-up content). python tests/test_adapters.py # or: pytest tests/ 1. call sequence: a recording fake stream checks the exact push_words / endpoint calls each adapter makes; 2. end to end: the same messages drive a real ArcTagger stream; checks words, speakers, utterances and JSON-able output. """ import json, os, sys HERE = os.path.dirname(os.path.abspath(__file__)); ROOT = os.path.dirname(HERE); sys.path.insert(0, ROOT) from arctag.asr import DeepgramAdapter, SpeechmaticsAdapter, summarize # noqa: E402 FIX = os.path.join(HERE, "fixtures") U1 = ["Hi,", "this", "is", "Dana", "from", "Acme", "Telecom."] U2 = ["How", "can", "I", "help", "you", "today?"] U3 = ["Yeah,", "my", "name", "is", "John", "Patel."] U4 = ["And", "my", "number", "is", "555-213-4489."] U5 = ["Thanks,", "John.", "I'll", "send", "a", "text", "to", "that", "number."] ALL = U1 + U2 + U3 + U4 + U5 def P(words, final, k): return ("push_words", {"words": words, "final": final, "revise_from": k}) E = ("endpoint", {}) EXPECT = { "deepgram": [P(["hi", "this", "is"], False, 0), P(["hi", "this", "is", "dana", "from"], False, 0), P(U1, True, 0), P(["how", "can", "i"], False, 7), P(U2, True, 7), E, P(["yeah", "my", "name"], False, 13), P(["yeah", "my", "name", "is", "john"], False, 13), P(U3, True, 13), P(["and", "my", "number", "is", "five", "five", "five"], False, 19), P(U4, True, 19), E, # UtteranceEnd arrived with interim words pending: endpoint after the final P(["thanks", "john"], False, 24), P(U5, True, 24), E], "speechmatics": [P(["Hi", "this"], False, 0), P(["Hi,", "this", "is", "Dana"], False, 0), P(U1, True, 0), P(["How", "can", "I"], False, 7), P(U2, True, 7), E, P(["Yeah", "my", "name"], False, 13), P(U3, True, 13), P(["And", "my", "number", "is", "555"], False, 19), P(U4, True, 19), E, # EndOfUtterance arrived with partial words pending P(["Thanks", "John"], False, 24), P(U5, True, 24), E], } ADAPTERS = {"deepgram": (DeepgramAdapter, "deepgram_live.jsonl"), "speechmatics": (SpeechmaticsAdapter, "speechmatics_rt.jsonl")} SPK = {"deepgram": ("S0", "S1"), "speechmatics": ("S1", "S2")} def messages(name): return [json.loads(l) for l in open(os.path.join(FIX, ADAPTERS[name][1])) if l.strip()] class FakeStream: """records calls; keeps the word list like the real stream does""" def __init__(self): self.words = []; self.n_endpoints = 0; self.spk = [] def push_words(self, words, final=True, revise_from=None, endpoint=False): k = len(self.words) if revise_from is None else revise_from assert 0 <= k <= len(self.words) self.words = self.words[:k] + [dict(w, final=final) for w in words] return {"n_words": len(self.words), "spans": [], "events": [], "utterances": [], "running": {}} def endpoint(self): assert all(w["final"] for w in self.words), "endpoint with non-final words" self.n_endpoints += 1 return {"n_words": len(self.words), "spans": [], "events": [], "utterances": [{}], "running": {}} def redacted_text(self): return " ".join(w["w"] for w in self.words) def check_sequence(name): cls, _ = ADAPTERS[name] fs = FakeStream(); ad = cls(fs) for m in messages(name): ad(m) assert ad.calls == EXPECT[name], json.dumps(ad.calls, indent=0) assert [w["w"] for w in fs.words] == ALL and all(w["final"] for w in fs.words) spk = [w.get("spk") for w in fs.words] a, b = SPK[name] assert spk == [a] * 13 + [b] * 11 + [a] * 9, spk assert fs.n_endpoints == 3 return len(ad.calls) def check_end_to_end(name, tg): cls, _ = ADAPTERS[name] st = tg.stream(domain="talk"); ad = cls(st); lines = [] for m in messages(name): for up in ad(m): lines.append(json.dumps(summarize(up, st))) # must be JSON-serialisable res = st.finalize() assert res["words"] == ALL, res["words"] assert [(u["start"], u["end"]) for u in res["utterances"]] == [(0, 13), (13, 24), (24, 33)], res["utterances"] assert [w.get("spk") for w in st.s.words][:2] == [SPK[name][0]] * 2 return {"updates": len(lines), "redacted_text": res["redacted_text"], "spans": [(s["tag"], s["text"]) for s in res["spans"]], "acts": [u["act"] for u in res["utterances"]], "roles": [u["role"] for u in res["utterances"]]} def test_deepgram_sequence(): check_sequence("deepgram") def test_speechmatics_sequence(): check_sequence("speechmatics") def test_end_to_end(): from arctag import ArcTagger tg = ArcTagger.from_pretrained(ROOT) for name in ADAPTERS: check_end_to_end(name, tg) if __name__ == "__main__": for name in ADAPTERS: print(f"{name}: call sequence OK ({check_sequence(name)} calls)") import torch torch.set_num_threads(1) from arctag import ArcTagger tg = ArcTagger.from_pretrained(ROOT) for name in ADAPTERS: print(f"{name}: end to end OK", json.dumps(check_end_to_end(name, tg)))