Public Access
251 lines
8.8 KiB
C#
251 lines
8.8 KiB
C#
using MeetingAssistant.Speakers;
|
|
using MeetingAssistant.Transcription;
|
|
using Microsoft.Extensions.Logging;
|
|
|
|
namespace MeetingAssistant.Recording;
|
|
|
|
internal sealed class SpeakerAudioSampleCollector
|
|
{
|
|
private readonly object gate = new();
|
|
private readonly RollingAudioBuffer audioBuffer;
|
|
private readonly Dictionary<string, List<SpeakerAudioSample>> samplesBySpeaker = new(StringComparer.OrdinalIgnoreCase);
|
|
private readonly int maxSamplesPerSpeaker;
|
|
private readonly TimeSpan minimumUninterruptedSpeechDuration;
|
|
private readonly TimeSpan maximumSegmentGap;
|
|
private readonly ILogger? logger;
|
|
private PendingSpeakerSpan? pendingSpan;
|
|
|
|
public SpeakerAudioSampleCollector(TimeSpan bufferDuration, int maxSamplesPerSpeaker)
|
|
: this(
|
|
bufferDuration,
|
|
maxSamplesPerSpeaker,
|
|
TimeSpan.FromSeconds(30),
|
|
TimeSpan.FromSeconds(1),
|
|
logger: null)
|
|
{
|
|
}
|
|
|
|
public SpeakerAudioSampleCollector(
|
|
TimeSpan bufferDuration,
|
|
int maxSamplesPerSpeaker,
|
|
TimeSpan minimumUninterruptedSpeechDuration,
|
|
TimeSpan maximumSegmentGap,
|
|
ILogger? logger = null)
|
|
{
|
|
audioBuffer = new RollingAudioBuffer(bufferDuration);
|
|
this.maxSamplesPerSpeaker = Math.Max(1, maxSamplesPerSpeaker);
|
|
this.minimumUninterruptedSpeechDuration = minimumUninterruptedSpeechDuration > TimeSpan.Zero
|
|
? minimumUninterruptedSpeechDuration
|
|
: TimeSpan.Zero;
|
|
this.maximumSegmentGap = maximumSegmentGap >= TimeSpan.Zero
|
|
? maximumSegmentGap
|
|
: TimeSpan.Zero;
|
|
this.logger = logger;
|
|
}
|
|
|
|
public void AppendAudio(AudioChunk chunk)
|
|
{
|
|
audioBuffer.Append(chunk);
|
|
}
|
|
|
|
public void Reset()
|
|
{
|
|
lock (gate)
|
|
{
|
|
samplesBySpeaker.Clear();
|
|
pendingSpan = null;
|
|
audioBuffer.Reset();
|
|
}
|
|
}
|
|
|
|
public SpeakerAudioSample? TryAdd(TranscriptionSegment segment)
|
|
{
|
|
if (!IsDiarizedSpeaker(segment.Speaker))
|
|
{
|
|
logger?.LogInformation(
|
|
"Discarding speaker identity sample for {Speaker} because the segment has no diarized speaker label",
|
|
segment.Speaker);
|
|
return null;
|
|
}
|
|
|
|
TranscriptionSegment sampleSegment;
|
|
PendingSpanReset? reset;
|
|
lock (gate)
|
|
{
|
|
(sampleSegment, reset) = ExtendPendingSpan(segment);
|
|
}
|
|
|
|
if (reset is not null)
|
|
{
|
|
logger?.LogInformation(
|
|
"Reset speaker identity sample span from {PreviousSpeaker} to {Speaker}: previous end {PreviousEnd}, next start {NextStart}, gap {Gap}, maximum gap {MaximumGap}",
|
|
reset.PreviousSpeaker,
|
|
segment.Speaker,
|
|
reset.PreviousEnd,
|
|
segment.Start,
|
|
reset.Gap,
|
|
maximumSegmentGap);
|
|
}
|
|
|
|
var score = Score(sampleSegment, minimumUninterruptedSpeechDuration);
|
|
if (!score.Accepted)
|
|
{
|
|
logger?.LogInformation(
|
|
"Discarding speaker identity sample for {Speaker} because {Reason}: duration {Duration}, minimum duration {MinimumDuration}, word count {WordCount}",
|
|
sampleSegment.Speaker,
|
|
score.Reason,
|
|
sampleSegment.End - sampleSegment.Start,
|
|
minimumUninterruptedSpeechDuration,
|
|
score.WordCount);
|
|
return null;
|
|
}
|
|
|
|
var wavBytes = audioBuffer.TryExtractWav(sampleSegment.Start, sampleSegment.End);
|
|
if (wavBytes.Length == 0)
|
|
{
|
|
logger?.LogInformation(
|
|
"Discarding speaker identity sample for {Speaker} because no audio could be extracted from rolling buffer: start {Start}, end {End}, duration {Duration}",
|
|
sampleSegment.Speaker,
|
|
sampleSegment.Start,
|
|
sampleSegment.End,
|
|
sampleSegment.End - sampleSegment.Start);
|
|
return null;
|
|
}
|
|
|
|
var sample = new SpeakerAudioSample(sampleSegment.Speaker, sampleSegment, wavBytes, score.Value);
|
|
lock (gate)
|
|
{
|
|
if (!samplesBySpeaker.TryGetValue(sampleSegment.Speaker, out var samples))
|
|
{
|
|
samples = [];
|
|
samplesBySpeaker[sampleSegment.Speaker] = samples;
|
|
}
|
|
|
|
samples.Add(sample);
|
|
var beforeCount = samples.Count;
|
|
var bestSamples = samples
|
|
.OrderByDescending(candidate => candidate.Score)
|
|
.Take(maxSamplesPerSpeaker)
|
|
.ToList();
|
|
var retained = bestSamples.Contains(sample);
|
|
samples.Clear();
|
|
samples.AddRange(bestSamples);
|
|
if (!retained)
|
|
{
|
|
logger?.LogInformation(
|
|
"Discarding speaker identity sample for {Speaker} because it was not among the best {MaxSamplesPerSpeaker} retained sample(s): score {Score}, candidate count {CandidateCount}",
|
|
sample.Speaker,
|
|
maxSamplesPerSpeaker,
|
|
sample.Score,
|
|
beforeCount);
|
|
return null;
|
|
}
|
|
}
|
|
|
|
return sample;
|
|
}
|
|
|
|
public IReadOnlyList<SpeakerAudioSample> Snapshot()
|
|
{
|
|
lock (gate)
|
|
{
|
|
return samplesBySpeaker.Values
|
|
.SelectMany(samples => samples)
|
|
.OrderBy(sample => sample.Speaker, StringComparer.OrdinalIgnoreCase)
|
|
.ThenByDescending(sample => sample.Score)
|
|
.ToList();
|
|
}
|
|
}
|
|
|
|
private static bool IsDiarizedSpeaker(string speaker)
|
|
{
|
|
return !string.IsNullOrWhiteSpace(speaker) &&
|
|
!string.Equals(speaker, "Unknown", StringComparison.OrdinalIgnoreCase);
|
|
}
|
|
|
|
private (TranscriptionSegment Segment, PendingSpanReset? Reset) ExtendPendingSpan(TranscriptionSegment segment)
|
|
{
|
|
if (pendingSpan is null ||
|
|
!SpeakerSampleSpanSelector.CanExtend(pendingSpan.Speaker, pendingSpan.End, segment, maximumSegmentGap))
|
|
{
|
|
var reset = pendingSpan is null
|
|
? null
|
|
: new PendingSpanReset(
|
|
pendingSpan.Speaker,
|
|
pendingSpan.End,
|
|
segment.Start - pendingSpan.End);
|
|
pendingSpan = new PendingSpeakerSpan(
|
|
segment.Speaker,
|
|
segment.Start,
|
|
segment.End,
|
|
[segment.Text]);
|
|
return (pendingSpan.ToSegment(), reset);
|
|
}
|
|
|
|
pendingSpan = pendingSpan.Extend(segment);
|
|
return (pendingSpan.ToSegment(), null);
|
|
}
|
|
|
|
private static SampleScore Score(
|
|
TranscriptionSegment segment,
|
|
TimeSpan minimumUninterruptedSpeechDuration)
|
|
{
|
|
var durationSeconds = (segment.End - segment.Start).TotalSeconds;
|
|
if (durationSeconds < minimumUninterruptedSpeechDuration.TotalSeconds)
|
|
{
|
|
return new SampleScore(false, 0, "speech duration is below the configured minimum", WordCount(segment.Text));
|
|
}
|
|
|
|
var words = WordCount(segment.Text);
|
|
if (words < 3)
|
|
{
|
|
return new SampleScore(false, 0, "word count is below the minimum useful sample length", words);
|
|
}
|
|
|
|
var durationScore = Math.Min(durationSeconds / Math.Max(1, minimumUninterruptedSpeechDuration.TotalSeconds), 2);
|
|
var wordScore = Math.Min(words / 60.0, 1);
|
|
var sentenceBonus = segment.Text.TrimEnd().EndsWith('.') ||
|
|
segment.Text.TrimEnd().EndsWith('?') ||
|
|
segment.Text.TrimEnd().EndsWith('!')
|
|
? 5
|
|
: 0;
|
|
return new SampleScore(true, durationScore * 70 + wordScore * 30 + sentenceBonus, null, words);
|
|
}
|
|
|
|
private static int WordCount(string text)
|
|
{
|
|
return text
|
|
.Split(' ', StringSplitOptions.RemoveEmptyEntries | StringSplitOptions.TrimEntries)
|
|
.Length;
|
|
}
|
|
|
|
private sealed record SampleScore(bool Accepted, double Value, string? Reason, int WordCount);
|
|
|
|
private sealed record PendingSpanReset(string PreviousSpeaker, TimeSpan PreviousEnd, TimeSpan Gap);
|
|
|
|
private sealed record PendingSpeakerSpan(
|
|
string Speaker,
|
|
TimeSpan Start,
|
|
TimeSpan End,
|
|
IReadOnlyList<string> TextParts)
|
|
{
|
|
public PendingSpeakerSpan Extend(TranscriptionSegment segment)
|
|
{
|
|
return this with
|
|
{
|
|
End = segment.End > End ? segment.End : End,
|
|
TextParts = TextParts.Append(segment.Text).ToList()
|
|
};
|
|
}
|
|
|
|
public TranscriptionSegment ToSegment()
|
|
{
|
|
return new TranscriptionSegment(
|
|
Start,
|
|
End,
|
|
Speaker,
|
|
string.Join(' ', TextParts.Where(part => !string.IsNullOrWhiteSpace(part))));
|
|
}
|
|
}
|
|
}
|