Documented the transcription bitrate decisions and hardened its fallback

This commit is contained in:
Thorsten Sommer 2026-09-19 18:47:16 +02:00
parent edfb9a5819
commit 3d0761f243
Signed by untrusted user who does not match committer: tsommer
GPG Key ID: 371BBA77A02C0108
7 changed files with 35 additions and 5 deletions

View File

@ -582,6 +582,9 @@ CONFIG["SETTINGS"] = {}
-- Allowed values are: KBPS_32, KBPS_64, KBPS_128, KBPS_256
-- Higher bitrates improve transcription accuracy on noisy or quiet recordings, at the cost of
-- a larger upload to the transcription provider. KBPS_128 is recommended.
-- Please note: this bitrate applies whenever a recording has to be re-encoded. A file which already
-- is a single mono 48 kHz Opus track in a WebM container and stays below 25 MiB is forwarded to the
-- transcription provider unchanged, keeping the bitrate it was created with.
-- CONFIG["SETTINGS"]["DataApp.OpusBitrate"] = "KBPS_32"
--
-- Allow the user to change the Opus bitrate even though your organization set a default above:

View File

@ -327,6 +327,8 @@ public static class ConfigurationSelectDataFactory
public static IEnumerable<ConfigurationSelectData<TranscriptionOpusBitrate>> GetTranscriptionOpusBitrateData()
{
foreach (var bitrate in Enum.GetValues<TranscriptionOpusBitrate>())
{
yield return new(bitrate.GetName(), bitrate);
}
}
}

View File

@ -115,6 +115,13 @@ public sealed class DataApp(Expression<Func<Data, DataApp>>? configSelection = n
/// <summary>
/// The Opus bitrate used when normalizing uploaded audio/video for transcription.
/// </summary>
/// <remarks>
/// Every recording is re-encoded to mono Opus before it goes to the transcription provider. That
/// encoding used to be fixed at 32 kbps, which cost the transcription models entire quiet passages:
/// a greeting spoken softly at the start of a recording was simply missing from the transcript. The
/// same recording compared at 64 and 128 kbps came back complete, which is why 128 kbps is the
/// default here. Users trading accuracy for a smaller upload can still pick a lower bitrate.
/// </remarks>
public TranscriptionOpusBitrate OpusBitrate { get; set; } = ManagedConfiguration.Register(configSelection, n => n.OpusBitrate, TranscriptionOpusBitrate.KBPS_128);
/// <summary>

View File

@ -2,8 +2,13 @@ namespace AIStudio.Settings.DataModel;
public enum TranscriptionOpusBitrate
{
// The recommended bitrate is deliberately the member with the underlying value 0: when the
// settings file holds a value TolerantEnumConverter cannot read, it falls back to that member.
// Landing on the lowest bitrate there would silently reintroduce the very defect this setting
// exists to prevent -- transcripts losing what was said quietly.
KBPS_128 = 0,
KBPS_32,
KBPS_64,
KBPS_128,
KBPS_256,
}

View File

@ -21,6 +21,9 @@ public static class TranscriptionOpusBitrateExtensions
TranscriptionOpusBitrate.KBPS_64 => 64_000,
TranscriptionOpusBitrate.KBPS_128 => 128_000,
TranscriptionOpusBitrate.KBPS_256 => 256_000,
_ => 32_000,
// A value we do not know must never mean the lowest quality: that is how quiet passages
// went missing from transcripts in the first place. Fall back to the recommended bitrate.
_ => 128_000,
};
}

View File

@ -647,9 +647,14 @@ public sealed class MediaTranscriptionService(
MediaOperation operation,
bool updateImportState)
{
// The bitrate governs re-encoding, not every upload: the runtime hands a file through
// unchanged when it already is a single mono 48 kHz Opus track in a WebM container and small
// enough. Such a file keeps whatever bitrate it was made with, which is the better outcome --
// re-encoding it could only take quality away, never add any.
var opusBitrateBps = settingsManager.ConfigurationData.App.OpusBitrate.GetBitsPerSecond();
// The quick POST is intentionally not cancelled: losing its response could orphan a job
// whose ID the client never received. Cancellation is applied immediately after ownership.
var opusBitrateBps = settingsManager.ConfigurationData.App.OpusBitrate.GetBitsPerSecond();
var jobId = await rustService.StartMediaJobAsync(mediaPath, normalizedPath, opusBitrateBps, CancellationToken.None);
operation.JobId = jobId;

View File

@ -60,6 +60,11 @@ const OPUS_PRE_SKIP: u16 = 312;
const DEFAULT_MAX_PASS_THROUGH_BYTES: u64 = 25 * 1024 * 1024;
/// Default target bitrate for mono speech-oriented Opus output, used when a request omits it.
///
/// AI Studio always states a bitrate, so this default only covers requests which leave it out. It is
/// 128 kbps because the 32 kbps this encoder used to be fixed at cost transcription models whole
/// quiet passages: a softly spoken greeting at the start of a recording never reached the transcript,
/// while the same recording at 128 kbps came back complete.
const DEFAULT_OPUS_BITRATE_BPS: u32 = 128_000;
/// Bounded input block used for streaming resampling.