Chunk overlap + overlap-aware stitching
Chunks were contiguous (start = prev end) with a naïve offset-concat stitch — no overlap. That cut sentences at boundaries, denied the diarizer context at edges, and let one voice split across chunks (the MH/Unknown_0 problem). Now each ~150s body is sliced with a 15s margin on both sides ([bodyStart-15, bodyEnd+15]); the stitcher keeps a segment only in the chunk that owns its MIDPOINT (body region) and drops it from the neighbour's margin — so boundary-spanning speech is seen whole by the backend and kept exactly once. - SessionPackager.PlannedChunk gains bodyStart/bodyEnd; planChunks adds overlapSeconds. - TranscriptAssembler.ChunkResult carries body bounds (defaults keep-all → no-overlap behaviour preserved for existing callers); assemble dedups by midpoint-in-body. - TranscriptPipeline passes body bounds through. Complements (doesn't replace) the fragment-smoothing + reconciliation safety nets; this is the upstream fix. ~+20% backend audio per interior chunk. 63/63 XCTest (new: overlap window layout + boundary-segment dedup).
This commit is contained in:
Grant Gilliam
@@ -10,24 +10,35 @@ import AVFoundation
|
||||
enum SessionPackager {
|
||||
struct PlannedChunk: Equatable {
|
||||
let index: Int
|
||||
let start: Double // global seconds
|
||||
let end: Double
|
||||
let start: Double // sliced window start (global seconds, incl. overlap margin)
|
||||
let end: Double // sliced window end (incl. overlap margin)
|
||||
let bodyStart: Double // the region this chunk OWNS (no overlap) — for stitch dedup
|
||||
let bodyEnd: Double
|
||||
}
|
||||
|
||||
/// One chunk if short; otherwise even ~`chunkSeconds` windows.
|
||||
/// One chunk if short; otherwise ~`chunkSeconds` bodies, each sliced with an
|
||||
/// `overlapSeconds` margin on both sides. The margin gives the backend context at
|
||||
/// boundaries (so a sentence isn't cut and the diarizer attributes edge speech
|
||||
/// correctly and keeps a voice consistent across chunks); the stitcher keeps only
|
||||
/// each chunk's owned `body` region, deduping the overlap.
|
||||
static func planChunks(durationSec: Double,
|
||||
chunkSeconds: Double = 150,
|
||||
overlapSeconds: Double = 15,
|
||||
thresholdSec: Double = 180) -> [PlannedChunk] {
|
||||
guard durationSec > thresholdSec else {
|
||||
return [PlannedChunk(index: 0, start: 0, end: durationSec)]
|
||||
return [PlannedChunk(index: 0, start: 0, end: durationSec, bodyStart: 0, bodyEnd: durationSec)]
|
||||
}
|
||||
var chunks: [PlannedChunk] = []
|
||||
var start = 0.0
|
||||
var bodyStart = 0.0
|
||||
var index = 0
|
||||
while start < durationSec - 0.001 {
|
||||
let end = min(start + chunkSeconds, durationSec)
|
||||
chunks.append(PlannedChunk(index: index, start: start, end: end))
|
||||
start = end
|
||||
while bodyStart < durationSec - 0.001 {
|
||||
let bodyEnd = min(bodyStart + chunkSeconds, durationSec)
|
||||
chunks.append(PlannedChunk(
|
||||
index: index,
|
||||
start: max(0, bodyStart - overlapSeconds),
|
||||
end: min(durationSec, bodyEnd + overlapSeconds),
|
||||
bodyStart: bodyStart, bodyEnd: bodyEnd))
|
||||
bodyStart = bodyEnd
|
||||
index += 1
|
||||
}
|
||||
return chunks
|
||||
|
||||
@@ -5,8 +5,13 @@ import Foundation
|
||||
/// name, and fingerprints collected for the voiceprint store.
|
||||
enum TranscriptAssembler {
|
||||
struct ChunkResult {
|
||||
let chunkStart: Double // global seconds
|
||||
let chunkStart: Double // global seconds (the sliced window start)
|
||||
let response: LabelMergeResponse
|
||||
// The region this chunk OWNS; segments whose midpoint falls outside it are the
|
||||
// neighbour's (overlap margin) and are dropped here. Defaults keep everything
|
||||
// (no-overlap behaviour).
|
||||
var bodyStart: Double = -.greatestFiniteMagnitude
|
||||
var bodyEnd: Double = .greatestFiniteMagnitude
|
||||
}
|
||||
|
||||
struct Assembled {
|
||||
@@ -40,13 +45,16 @@ enum TranscriptAssembler {
|
||||
|
||||
for chunk in chunks {
|
||||
let offset = chunk.chunkStart
|
||||
// Audio length from the chunk window, so silent/all-unknown calls still
|
||||
// report a real duration (not just the last segment's end).
|
||||
duration = max(duration, offset + chunk.response.duration)
|
||||
// Body end bounds the real session length even on silent/all-unknown calls.
|
||||
duration = max(duration, min(chunk.bodyEnd, offset + chunk.response.duration))
|
||||
|
||||
for seg in chunk.response.segments {
|
||||
let start = seg.startSeconds + offset
|
||||
let end = seg.endSeconds + offset
|
||||
// Overlap dedup: keep a segment only in the chunk that OWNS its midpoint;
|
||||
// the other chunk saw it only in its margin (for context) and drops it.
|
||||
let mid = (start + end) / 2
|
||||
guard mid >= chunk.bodyStart, mid < chunk.bodyEnd else { continue }
|
||||
segments.append(.init(start: start, end: end, speaker: seg.speaker, text: seg.text))
|
||||
duration = max(duration, end)
|
||||
}
|
||||
|
||||
@@ -91,7 +91,8 @@ final class TranscriptPipeline {
|
||||
known[name] = fp
|
||||
}
|
||||
voiceprints.update(with: response)
|
||||
results.append(.init(chunkStart: chunk.start, response: response))
|
||||
results.append(.init(chunkStart: chunk.start, response: response,
|
||||
bodyStart: chunk.bodyStart, bodyEnd: chunk.bodyEnd))
|
||||
}
|
||||
await progress?(plan.count, plan.count)
|
||||
|
||||
|
||||
@@ -8,11 +8,33 @@ final class Phase5Tests: XCTestCase {
|
||||
XCTAssertEqual(c[0].end, 70, accuracy: 0.001)
|
||||
}
|
||||
|
||||
func testPlanChunksLong() {
|
||||
let c = SessionPackager.planChunks(durationSec: 400, chunkSeconds: 150)
|
||||
func testPlanChunksLongOverlapsWindowsWithContiguousBodies() {
|
||||
let c = SessionPackager.planChunks(durationSec: 400, chunkSeconds: 150, overlapSeconds: 15)
|
||||
XCTAssertEqual(c.count, 3)
|
||||
XCTAssertEqual(c[0].start, 0); XCTAssertEqual(c[0].end, 150)
|
||||
XCTAssertEqual(c[1].start, 150); XCTAssertEqual(c[2].end, 400)
|
||||
// Owned bodies tile the call with no gaps/overlap.
|
||||
XCTAssertEqual(c[0].bodyStart, 0); XCTAssertEqual(c[0].bodyEnd, 150)
|
||||
XCTAssertEqual(c[1].bodyStart, 150); XCTAssertEqual(c[1].bodyEnd, 300)
|
||||
XCTAssertEqual(c[2].bodyEnd, 400)
|
||||
// Sliced windows overlap by the margin (and clamp at the ends).
|
||||
XCTAssertEqual(c[0].start, 0); XCTAssertEqual(c[0].end, 165) // +15 trailing
|
||||
XCTAssertEqual(c[1].start, 135) // -15 leading
|
||||
XCTAssertLessThan(c[1].start, c[0].end) // windows overlap
|
||||
XCTAssertEqual(c[2].end, 400) // clamped
|
||||
}
|
||||
|
||||
func testAssembleDedupsOverlapByBody() {
|
||||
// A segment at global 152–156 sits in chunk1's body but also in chunk0's
|
||||
// trailing margin (overlap). It must be kept exactly once (by chunk1).
|
||||
let r0 = #"{"duration":165,"speakers":[{"cluster":"S0","name":"A","source":"visual","overlap_confidence":0.9}],"segments":[{"start_ms":152000,"end_ms":156000,"speaker":"A","text":"boundary"}],"fingerprints":{},"models":{}}"#
|
||||
let r1 = #"{"duration":180,"speakers":[{"cluster":"S0","name":"A","source":"visual","overlap_confidence":0.9}],"segments":[{"start_ms":17000,"end_ms":21000,"speaker":"A","text":"boundary"}],"fingerprints":{},"models":{}}"#
|
||||
let c0 = try! JSONDecoder().decode(LabelMergeResponse.self, from: Data(r0.utf8))
|
||||
let c1 = try! JSONDecoder().decode(LabelMergeResponse.self, from: Data(r1.utf8))
|
||||
let asm = TranscriptAssembler.assemble(sessionId: "s", app: "meet", chunks: [
|
||||
.init(chunkStart: 0, response: c0, bodyStart: 0, bodyEnd: 150),
|
||||
.init(chunkStart: 135, response: c1, bodyStart: 150, bodyEnd: 300),
|
||||
])
|
||||
XCTAssertEqual(asm.speakersFile.segments.count, 1) // deduped
|
||||
XCTAssertEqual(asm.speakersFile.segments[0].start, 152, accuracy: 0.01)
|
||||
}
|
||||
|
||||
func testRebaseClipsAndRebases() throws {
|
||||
|
||||
Reference in New Issue
Block a user