Meeting speaker diarization — feature spec & exploration#68
Open
gbrunoo wants to merge 5 commits into
Open
Conversation
Owner
Implements the spec from PR sebsto#64: splits the system-audio ("Others") track into per-speaker labels (Speaker 1, Speaker 2, …) using FluidAudio's SortformerDiarizer. The mic track stays "You". - MeetingSpeaker becomes .you / .others(speakerIndex: Int?) with a displayName computed property (0-based storage, +1 only at display). - New MeetingDiarizer actor wraps SortformerDiarizer: warmUp() loads the CoreML model, ingest() feeds chunks, dominantSpeaker(in:) resolves the speaker over a time window. Sortformer slots map to monotonic per-meeting indices. Failures degrade gracefully to plain "Others". - MeetingAudioEngine exposes a single sampleRate constant and emits the system-audio stream as timestamped SystemAudioChunk values. - MeetingStateManager warms up the diarizer on start (when enabled), routes system chunks through ingest + dominantSpeaker, resets on stop. - MeetingTranscriptView uses displayName with a per-speaker color palette. - SettingsStore gains meetingDiarizationEnabled (default false, opt-in); Settings adds a Meeting section toggle. - Cold-start nil speaker renders as "Others" until the diarizer warms up. Tests: MeetingDiarizerTests for the pre-init no-op contract, updated MeetingTranscript/AudioEngine tests for the new APIs. Diarization quality is verified via the manual E2E plan in the spec.
The diarizer warmUp() (model load/download, several seconds) ran before meetingState flipped to .recording, so the button stayed on 'Start Meeting' with no feedback, and the meetingState==.idle guard was still true during the gap — a second click double-triggered startCapture(). - Add an isStarting re-entrancy guard. - Start capture and flip to .recording immediately for instant feedback; warm up the diarizer concurrently inside the recording task group. - Check diarizationActive per-chunk so diarization engages once warmup completes mid-meeting. - MeetingDiarizer now tracks the meeting-relative time of its first ingested chunk and offsets segment times, keeping speaker windows aligned when the diarizer starts after the meeting has begun. - stopMeeting resets the diarizer unconditionally (safe no-op) to avoid a race with the concurrent warmup task.
- Switch Sortformer from the default fastV2_1 config to balancedV2: same ~1s latency but a much larger FIFO (188 vs 40) for better quality, and v2 weights handle overlapping/high-speaker-count audio better than v2.1 (which is documented to degrade and likely caused phantom speakers). - Give the system-audio path its own shorter chunk size (~3s vs the mic's ~5s) so each transcript entry spans less time, reducing how often a speaker change inside one entry gets flattened to a single label.
Local editor config and agent skills should not be committed to the repo.
gbrunoo
force-pushed
the
meeting-diarizer
branch
from
c206890 to
8bfbf52
Compare
Collaborator
Author
|
Synced now :) |
Fixes issue sebsto#65: without headphones, remote participants' speech plays out of the speakers and leaks back into the microphone, so the same utterance was transcribed twice — once as 'Others' (system audio) and once as 'You' (mic echo). - MeetingTranscript.appendSuppressingEcho() de-duplicates across the two tracks: a 'You' entry that closely matches a recent 'Others' entry is dropped, and an arriving 'Others' entry removes a prior 'You' echo (handles either arrival order). Matching uses normalized text + Levenshtein similarity within an 8s window, skipping short (<3 word) backchannels to avoid dropping genuine speech. - MeetingStateManager routes both transcription paths through record(), which applies suppression when enabled (text is never logged). - SettingsStore gains meetingEchoSuppressionEnabled (default true, opt-out); Settings adds a 'Suppress Microphone Echo' toggle. - Tests: MeetingEchoSuppressionTests for the de-dup contract and SettingsStore default/persistence/restore coverage.
Collaborator
Author
|
@sebsto I might have found a fix for #65 but it is clearly a beta feature for now. Approach: since true audio-level AEC isn't feasible for audio played by another app, this de-duplicates at the transcript layer. A mic ("You") entry is compared against the concatenation of nearby "Others" speech and suppressed when ≥60% of its words are covered (fuzzy per-word matching to absorb ASR drift, e.g. "démonté"/"démonter"). It handles both arrival orders, mic-first entries are pruned once the matching remote lines come in. It's a best-effort heuristic, so I've labeled the Settings toggle "Suppress Microphone Echo (Beta)" (default on, opt-out). Two tunable knobs: Known limitation: in the mic-first case there's a brief visible flicker ("You" line appears, then is replaced by the diarized remote line). Happy to add a short grace delay to remove it if you'd prefer. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Branch setup for exploring speaker diarization in the meeting transcriber — separating the "Others" audio stream into individual speakers ("Speaker 1", "Speaker 2", etc.).
Branched from the meeting-transcriber work (#63) with the upstream SPM layout (#62) already merged in, so this stays in sync with
main.Background
The current meeting transcriber does simple binary routing:
This is attribution by input source, not by voice identity. Diarization would identify distinct speakers within the system audio stream.
Proposed approach
MeetingTranscript.Speakerfrom a fixed enum to dynamic speaker identities (e.g..other(id: Int, label: String?))Constraints to respect
Status
🚧 Exploratory / spec stage — opening this against
spec/meeting-speaker-diarizationto align on design before implementation.Open questions