Skip to content

fix: skip firstTokenLogProbThreshold when promptTokens are set #438

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
alan890104 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix: skip firstTokenLogProbThreshold when promptTokens are set #438

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5606dea
fix: skip firstTokenLogProbThreshold when promptTokens are set
alan890104 Mar 11, 2026
6dafa8a
test: add regression test for promptTokens with firstTokenLogProbThre…
alan890104 Mar 11, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion Sources/WhisperKit/Core/TextDecoder.swift
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -850,7 +850,7 @@ open class TextDecoder: TextDecoding, WhisperMLModel {
timings.decodingSampling += samplingTime

isFirstTokenLogProbTooLow =
if isFirstToken, let firstTokenLogProbThreshold = options.firstTokenLogProbThreshold, nextTokenLogProb < firstTokenLogProbThreshold {
if isFirstToken, options.promptTokens == nil, let firstTokenLogProbThreshold = options.firstTokenLogProbThreshold, nextTokenLogProb < firstTokenLogProbThreshold {
true
} else {
Comment on lines 852 to 855
Copy link

Copilot AI Mar 12, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This change makes firstTokenLogProbThreshold effectively a no-op whenever options.promptTokens is non-nil. That’s a behavior change for an existing public option, so it should be documented (e.g., in DecodingOptions docs / README) to avoid confusing callers who set a strict threshold expecting it to be enforced.

Copilot uses AI. Check for mistakes.
false
Expand Down
33 changes: 33 additions & 0 deletions Tests/WhisperKitTests/UnitTests.swift
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1702,6 +1702,39 @@ final class UnitTests: XCTestCase {
XCTAssertFalse(result.text.contains(promptText), "Prompt text should not be present in the result")
}

func testPromptTokensWithStrictFirstTokenThreshold() async throws {
// Regression test for https://github.com/argmaxinc/WhisperKit/issues/372
//
// promptTokens shift the decoder's first content token logprob downward.
// Measured on tiny model + jfk.wav:
// - Without prompt: firstToken logprob ≈ -0.087
// - With CJK prompt: firstToken logprob ≈ -0.578
//
// On larger/turbo models with longer audio, this shift is amplified
// enough to breach the default threshold (-1.5), causing empty output.
// We use -0.5 here to reliably reproduce the issue on the tiny model,
// simulating the larger shift seen on turbo variants.
let config = WhisperKitConfig(model: "tiny", verbose: true, logLevel: .debug, load: true)
let whisperKit = try await WhisperKit(config)
let tokenizer = try XCTUnwrap(whisperKit.tokenizer)

let promptText = "繁體中文語音轉錄。"
let promptTokens = tokenizer.encode(text: promptText)

let options = DecodingOptions(
skipSpecialTokens: true,
promptTokens: promptTokens,
firstTokenLogProbThreshold: -0.5
)

let result = try await XCTUnwrapAsync(
await transcribe(with: .tiny, options: options),
"Failed to transcribe"
)

XCTAssertFalse(result.text.isEmpty, "Transcription should not be empty when promptTokens are set, even with a strict firstTokenLogProbThreshold")
}

func testPrefixTokens() async throws {
let config = WhisperKitConfig(model: "tiny", verbose: true, logLevel: .debug, load: true)
let whisperKit = try await WhisperKit(config)
Expand Down
Loading