Grant Gilliam
1c133c8970
Investigating Grant's real 38-min group call: 'Marty' was a GARBAGE cluster (192 segs, 0.37s mean, 186 ≤2 words, 125 single words flanked by the same other speaker — diarization micro-fragments split mid-sentence, then LLM-named 'Marty'). Same for 'Message'/'HI'. - SpeakerReconciler.smoothFragments: dissolve non-self clusters whose MEDIAN segment duration ≤ 1s (≥3 segs) — reassign each fragment to the temporally-nearest real speaker. (Median, not max, so one stray long segment can't rescue a fragment cluster — the bug in the first cut.) On the real call: 7 speakers (3 junk) → 4 real (Marty/Message/HI absorbed into Grant/Jonathan/Me/MH). Runs before LLM naming. - LLM naming guardrails: forbid assigning the self name or ANY already-taken name to another voice (fixes 'Grant' = the user's name pinned on a remote speaker); prompt demands self-intro / direct-address evidence (mention ≠ presence), 'precision over coverage', one name per speaker. - Open saved session now offers Open Editor vs Re-process, so newer logic can be applied to past calls (+ always-visible progress from the prior fix). NOTE: the self-name guardrail needs the app to KNOW the user's name — selfName is still 'Me', so set it in Settings (e.g. 'Grant') so the LLM can't reuse it. 62/62 XCTest.
84 lines
4.9 KiB
Swift
84 lines
4.9 KiB
Swift
import XCTest
|
|
@testable import Ten31Transcripts
|
|
|
|
final class SpeakerReconcilerTests: XCTestCase {
|
|
private func seg(_ s: Double, _ e: Double, _ who: String) -> SpeakersFile.Segment {
|
|
.init(start: s, end: e, speaker: who, text: "x")
|
|
}
|
|
private func file(_ speakers: [SpeakersFile.Speaker], _ segs: [SpeakersFile.Segment]) -> SpeakersFile {
|
|
SpeakersFile(sessionId: "s", app: "signal", durationSec: 60, speakers: speakers, segments: segs, models: [:])
|
|
}
|
|
private func sp(_ name: String, _ source: String, _ overlap: Double? = nil) -> SpeakersFile.Speaker {
|
|
.init(name: name, source: source, overlapConfidence: overlap, matchSimilarity: nil)
|
|
}
|
|
|
|
func testCosine() {
|
|
XCTAssertEqual(SpeakerReconciler.cosine([1, 0, 0], [1, 0, 0]), 1.0, accuracy: 0.0001)
|
|
XCTAssertEqual(SpeakerReconciler.cosine([1, 0, 0], [0, 1, 0]), 0.0, accuracy: 0.0001)
|
|
XCTAssertEqual(SpeakerReconciler.cosine([1, 0], [3, 0]), 1.0, accuracy: 0.0001) // scale-invariant
|
|
}
|
|
|
|
func testMergeCollapsesSplitPersonIntoNamedSurvivor() {
|
|
// The 1-on-1 failure: a remote person split into "MH" (named) + "Unknown_0".
|
|
let f = file([sp("Grant", "mic_channel"), sp("MH", "visual", 0.17), sp("Unknown_0", "unmatched")],
|
|
[seg(0, 2, "Grant"), seg(2, 4, "MH"), seg(4, 6, "Unknown_0"), seg(6, 8, "MH")])
|
|
let fps = ["MH": [Float(1), 0, 0], "Unknown_0": [Float(0.99), 0.02, 0], "Grant": [Float(0), 1, 0]]
|
|
let out = SpeakerReconciler.mergeByFingerprint(f, fingerprints: fps,
|
|
protected: SpeakerReconciler.protectedNames(f, selfName: "Grant"), threshold: 0.9)
|
|
XCTAssertEqual(Set(out.speakers.map { $0.name }), ["Grant", "MH"]) // Unknown_0 absorbed
|
|
XCTAssertFalse(out.segments.contains { $0.speaker == "Unknown_0" })
|
|
XCTAssertEqual(out.segments.filter { $0.speaker == "MH" }.count, 3) // 2 MH + 1 ex-Unknown
|
|
XCTAssertEqual(out.segments.filter { $0.speaker == "Grant" }.count, 1) // self untouched
|
|
}
|
|
|
|
func testNeverMergesSelf() {
|
|
// Self (mic_channel) must not merge even with an identical fingerprint.
|
|
let f = file([sp("Grant", "mic_channel"), sp("MH", "visual"), sp("JK", "visual")],
|
|
[seg(0, 2, "Grant"), seg(2, 4, "MH"), seg(4, 6, "JK")])
|
|
let fps = ["Grant": [Float(1), 0, 0], "MH": [Float(1), 0, 0], "JK": [Float(0), 1, 0]]
|
|
let out = SpeakerReconciler.mergeByFingerprint(f, fingerprints: fps,
|
|
protected: SpeakerReconciler.protectedNames(f, selfName: "Grant"), threshold: 0.9)
|
|
XCTAssertTrue(out.speakers.contains { $0.name == "Grant" })
|
|
XCTAssertEqual(out.speakers.count, 3) // nothing merged (MH/JK distinct, Grant protected)
|
|
}
|
|
|
|
func testKeepsDistinctVoices() {
|
|
let f = file([sp("MH", "visual"), sp("JK", "visual")], [seg(0, 2, "MH"), seg(2, 4, "JK")])
|
|
let fps = ["MH": [Float(1), 0, 0], "JK": [Float(0), 1, 0]]
|
|
let out = SpeakerReconciler.mergeByFingerprint(f, fingerprints: fps, protected: [], threshold: 0.8)
|
|
XCTAssertEqual(out.speakers.count, 2)
|
|
}
|
|
|
|
func testSmoothDissolvesFragmentCluster() {
|
|
// "Frag" is mostly micro-segments (the Marty pattern: median ≤ 1s) even though
|
|
// it has one longer stray → still absorbed into the surrounding real speaker.
|
|
let f = file([sp("Grant", "content"), sp("Frag", "content")],
|
|
[seg(0, 4, "Grant"), seg(4.0, 4.3, "Frag"), seg(4.4, 8, "Grant"),
|
|
seg(20, 24, "Grant"), seg(24.0, 24.2, "Frag"), seg(24.3, 28, "Grant"),
|
|
seg(30, 30.3, "Frag"), seg(31, 33, "Frag")]) // 4 Frag: 3 micro + 1 stray 2s
|
|
let out = SpeakerReconciler.smoothFragments(f, protected: [])
|
|
XCTAssertEqual(Set(out.speakers.map { $0.name }), ["Grant"]) // median(Frag)=0.3 ≤1 → dissolved
|
|
XCTAssertFalse(out.segments.contains { $0.speaker == "Frag" })
|
|
}
|
|
|
|
func testSmoothKeepsRealSpeakerWithMostlyLongSegs() {
|
|
let f = file([sp("A", "content")], [seg(0, 3, "A"), seg(3, 6, "A"), seg(6, 6.2, "A")]) // median 3 → real
|
|
XCTAssertEqual(SpeakerReconciler.smoothFragments(f, protected: []).speakers.map { $0.name }, ["A"])
|
|
}
|
|
|
|
func testSmoothProtectsSelfEvenIfAllShort() {
|
|
let f = file([sp("Me", "mic_channel"), sp("A", "content")],
|
|
[seg(0, 0.3, "Me"), seg(1, 4, "A"), seg(4, 4.2, "Me")])
|
|
let out = SpeakerReconciler.smoothFragments(f, protected: ["Me"])
|
|
XCTAssertTrue(out.speakers.contains { $0.name == "Me" }) // self never dissolved
|
|
}
|
|
|
|
func testParseNamingDropsNullAndKeepsConfidence() {
|
|
let json = #"{"speakers":[{"current":"MH","name":"Jonathan Kirkwood","confidence":"high"},{"current":"Unknown_0","name":null,"confidence":"low"}]}"#
|
|
let m = SpeakerReconciler.parseNaming(json)
|
|
XCTAssertEqual(m["MH"]?.name, "Jonathan Kirkwood")
|
|
XCTAssertEqual(m["MH"]?.confidence, "high")
|
|
XCTAssertNil(m["Unknown_0"]) // null name → not a rename
|
|
}
|
|
}
|