aria_kit/transcription/
mod.rs

1//! The speech-to-text contract engine crates implement.
2
3mod agreement;
4pub mod layer;
5mod native;
6mod session;
7mod spacing;
8
9pub use layer::{Layer, Layered, RemoveHesitations, Vocabulary};
10pub use spacing::Caret;
11
12use std::path::PathBuf;
13
14#[cfg(not(target_family = "wasm"))]
15use crate::audio::RecordedAudio;
16use crate::audio::{CANONICAL_SAMPLE_RATE, CaptureError, RecordingReader};
17use crate::backend::{BackendError, ModelError};
18use crate::cancel::CancelToken;
19pub use agreement::Cadence;
20pub use session::Session;
21
22/// The embedder-facing surface, re-exported by `aria-stt` so a consumer
23/// imports one library. Expert knobs — cancellation, event
24/// delivery, custom cadences — stay out and take an explicit `aria-kit`
25/// dependency.
26pub mod prelude {
27    #[cfg(all(feature = "silero", not(target_family = "wasm")))]
28    pub use super::session::HeardSession;
29    pub use super::{
30        Caret, Language, Layer, RemoveHesitations, Session, Streaming, Transcript,
31        TranscriptSegment, Transcription, TranscriptionError, Vocabulary, Whole,
32    };
33    pub use crate::audio::RecordingReader;
34    #[cfg(all(feature = "microphone", not(target_family = "wasm")))]
35    pub use crate::audio::{Microphone, RecordingMode};
36    pub use crate::cancel::CancelToken;
37    pub use crate::speaker::{Diarizer, Embedder};
38    #[cfg(all(feature = "silero", not(target_family = "wasm")))]
39    pub use crate::speaker::{Heard, SegmentLabeler, SegmentReader, Verdict};
40}
41
42const MAX_LANGUAGE_CODE_BYTES: usize = 64;
43
44/// The contract every speech-to-text engine implements.
45///
46/// An implementation must:
47/// - poll `cancel` at its natural pause points and answer
48///   [`TranscriptionError::Cancelled`];
49/// - accept only its supported [`TranscriptionAudio`] representation;
50/// - wrap its own failures in [`BackendError::engine`] under
51///   [`TranscriptionError::Load`] or [`TranscriptionError::Inference`];
52/// - refuse languages it cannot honour — [`english_only`] covers
53///   English-only models;
54/// - validate model paths with [`validate_model_file`] or
55///   [`validate_model_directory`] before reading them, on targets that have
56///   a filesystem.
57pub trait Transcriber: Send + Sync + 'static {
58    fn validate_language(
59        &self,
60        language: &Language,
61        cancel: &CancelToken,
62    ) -> Result<(), TranscriptionError>;
63
64    /// Transcribes one bounded audio input: a whole recording or a streaming
65    /// buffer, which the engine cannot tell apart.
66    fn transcribe_with(
67        &self,
68        audio: TranscriptionAudio<'_>,
69        language: &Language,
70        cancel: &CancelToken,
71    ) -> Result<Hypothesis, TranscriptionError>;
72
73    /// A decoder with its own state that takes audio a chunk at a time, for
74    /// an engine whose encoder streams. `None`, the default, means the
75    /// engine decodes bounded inputs only, and streaming runs by agreement.
76    /// A wrapper around another transcriber must forward this, or a native
77    /// stream fails with [`TranscriptionError::NoNativeStream`].
78    fn open_stream(
79        &self,
80        language: &Language,
81        cancel: &CancelToken,
82    ) -> Result<Option<Box<dyn ChunkStream>>, TranscriptionError> {
83        let _ = (language, cancel);
84        Ok(None)
85    }
86}
87
88/// One live decode of an engine that streams natively: audio in by the
89/// chunk, final words out. A word that came out is never revised.
90pub trait ChunkStream: Send {
91    /// How many canonical samples the encoder takes per step.
92    fn chunk_samples(&self) -> usize;
93
94    /// Feeds one chunk, exactly [`chunk_samples`](Self::chunk_samples)
95    /// long, and answers with what it let go of.
96    fn feed(&mut self, chunk: &[f32]) -> Result<Emitted, TranscriptionError>;
97}
98
99/// What one chunk let go of. A model writes words in pieces, and a word can
100/// finish in the chunk after it started; `joins_previous` says the first
101/// piece here continues the previous chunk's last word rather than opening
102/// a new one. Word timings are seconds from the start of the chunk.
103pub struct Emitted {
104    pub words: Hypothesis,
105    pub joins_previous: bool,
106}
107
108/// How an engine streams: by re-decoding a growing buffer until two
109/// decodes agree on its words, or natively, one chunk at a time.
110#[derive(Clone, Copy)]
111pub enum Stream {
112    Agreement(Cadence),
113    Native,
114}
115
116/// Shared construction contract: an engine name, configuration, validation, and loading.
117pub trait Engine: Transcriber + Sized {
118    const NAME: &'static str;
119
120    /// The one audio representation this engine accepts. Callers read it to
121    /// hand over the right input instead of guessing and failing at runtime.
122    const INPUT: AudioInputKind;
123
124    type Config;
125
126    /// Resolves what a load would read, without reading it or starting a
127    /// child.
128    fn check_from(config: &Self::Config) -> Result<PathBuf, TranscriptionError>;
129
130    fn load_from(config: Self::Config, cancel: &CancelToken) -> Result<Self, TranscriptionError>;
131}
132
133/// An engine with a tested streaming cadence or native chunking policy.
134pub trait StreamingEngine: Engine {
135    const STREAM: Stream;
136}
137
138impl<T: Transcriber + ?Sized> Transcriber for std::sync::Arc<T> {
139    fn validate_language(
140        &self,
141        language: &Language,
142        cancel: &CancelToken,
143    ) -> Result<(), TranscriptionError> {
144        (**self).validate_language(language, cancel)
145    }
146
147    fn transcribe_with(
148        &self,
149        audio: TranscriptionAudio<'_>,
150        language: &Language,
151        cancel: &CancelToken,
152    ) -> Result<Hypothesis, TranscriptionError> {
153        (**self).transcribe_with(audio, language, cancel)
154    }
155
156    fn open_stream(
157        &self,
158        language: &Language,
159        cancel: &CancelToken,
160    ) -> Result<Option<Box<dyn ChunkStream>>, TranscriptionError> {
161        (**self).open_stream(language, cancel)
162    }
163}
164
165impl<T: Transcriber + ?Sized> Transcriber for Box<T> {
166    fn validate_language(
167        &self,
168        language: &Language,
169        cancel: &CancelToken,
170    ) -> Result<(), TranscriptionError> {
171        (**self).validate_language(language, cancel)
172    }
173
174    fn transcribe_with(
175        &self,
176        audio: TranscriptionAudio<'_>,
177        language: &Language,
178        cancel: &CancelToken,
179    ) -> Result<Hypothesis, TranscriptionError> {
180        (**self).transcribe_with(audio, language, cancel)
181    }
182
183    fn open_stream(
184        &self,
185        language: &Language,
186        cancel: &CancelToken,
187    ) -> Result<Option<Box<dyn ChunkStream>>, TranscriptionError> {
188        (**self).open_stream(language, cancel)
189    }
190}
191
192#[derive(Debug, Clone, Copy, PartialEq, Eq)]
193pub enum AudioInputKind {
194    Samples,
195    Wav,
196}
197
198impl std::fmt::Display for AudioInputKind {
199    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
200        match self {
201            Self::Samples => formatter.write_str("canonical samples"),
202            Self::Wav => formatter.write_str("a completed WAV"),
203        }
204    }
205}
206
207/// One bounded input to a speech-to-text engine.
208pub enum TranscriptionAudio<'a> {
209    Samples(&'a [f32]),
210    #[cfg(not(target_family = "wasm"))]
211    Wav(&'a RecordedAudio),
212}
213
214impl<'a> TranscriptionAudio<'a> {
215    pub fn kind(&self) -> AudioInputKind {
216        match self {
217            Self::Samples(_) => AudioInputKind::Samples,
218            #[cfg(not(target_family = "wasm"))]
219            Self::Wav(_) => AudioInputKind::Wav,
220        }
221    }
222
223    pub fn require_samples(self, backend: &'static str) -> Result<&'a [f32], TranscriptionError> {
224        #[cfg(target_family = "wasm")]
225        let _ = backend;
226        match self {
227            Self::Samples(samples) => Ok(samples),
228            #[cfg(not(target_family = "wasm"))]
229            Self::Wav(_) => Err(TranscriptionError::UnsupportedAudioInput {
230                backend,
231                received: AudioInputKind::Wav,
232                required: AudioInputKind::Samples,
233            }),
234        }
235    }
236
237    #[cfg(not(target_family = "wasm"))]
238    pub fn require_wav(
239        self,
240        backend: &'static str,
241    ) -> Result<&'a RecordedAudio, TranscriptionError> {
242        match self {
243            Self::Wav(audio) => Ok(audio),
244            Self::Samples(_) => Err(TranscriptionError::UnsupportedAudioInput {
245                backend,
246                received: AudioInputKind::Samples,
247                required: AudioInputKind::Wav,
248            }),
249        }
250    }
251}
252
253impl std::fmt::Debug for TranscriptionAudio<'_> {
254    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
255        formatter
256            .debug_struct("TranscriptionAudio")
257            .field("kind", &self.kind())
258            .finish_non_exhaustive()
259    }
260}
261
262/// One decoded word's time inside the decoded audio window, in seconds.
263#[derive(Debug, Clone, Copy, PartialEq)]
264pub struct TimedWord {
265    pub start: f32,
266    pub end: f32,
267}
268
269impl TimedWord {
270    /// Where the word sits in the stream, for a window opening at `origin`.
271    pub fn span_from(&self, origin: u64) -> std::ops::Range<u64> {
272        let position = |seconds: f32| {
273            let samples = (f64::from(seconds) * f64::from(CANONICAL_SAMPLE_RATE)).round() as u64;
274            origin.saturating_add(samples)
275        };
276        let start = position(self.start);
277        start..position(self.end).max(start)
278    }
279}
280
281/// Text decoded from one bounded audio input. Silence is empty.
282pub struct Hypothesis {
283    text: String,
284    words: Vec<TimedWord>,
285}
286
287impl Hypothesis {
288    pub fn new(text: impl Into<String>) -> Self {
289        Self {
290            text: text.into().trim().to_owned(),
291            words: Vec::new(),
292        }
293    }
294
295    /// Timings aligned one-to-one with the text's whitespace words. Dropped
296    /// whole when the counts disagree — misaligned timings never escape.
297    pub fn with_words(text: impl Into<String>, words: Vec<TimedWord>) -> Self {
298        let mut hypothesis = Self::new(text);
299        if hypothesis.text.split_whitespace().count() == words.len() {
300            hypothesis.words = words;
301        }
302        hypothesis
303    }
304
305    pub fn as_str(&self) -> &str {
306        &self.text
307    }
308
309    /// Empty when the text has no aligned timings.
310    pub fn words(&self) -> &[TimedWord] {
311        &self.words
312    }
313
314    pub fn into_text(self) -> String {
315        self.text
316    }
317}
318
319/// Times the words the tokens spell. Each token is its text and its time
320/// inside the window. A word runs from one non-blank to the next blank;
321/// it starts with its first token and ends with its last. U+2581,
322/// SentencePiece's marker, counts as a blank, so the words match what
323/// [`Hypothesis::with_words`] counts once the text spells that marker as
324/// a space.
325pub fn word_timings<'a>(tokens: impl IntoIterator<Item = (&'a str, TimedWord)>) -> Vec<TimedWord> {
326    let mut words: Vec<TimedWord> = Vec::new();
327    let mut open = false;
328    for (text, timing) in tokens {
329        for character in text.chars() {
330            if character.is_whitespace() || character == '\u{2581}' {
331                open = false;
332                continue;
333            }
334            match words.last_mut() {
335                Some(word) if open => word.end = word.end.max(timing.end),
336                _ => {
337                    words.push(timing);
338                    open = true;
339                }
340            }
341        }
342    }
343    words
344}
345
346impl std::fmt::Debug for Hypothesis {
347    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
348        formatter
349            .debug_struct("Hypothesis")
350            .field("bytes", &self.text.len())
351            .field("timed_words", &self.words.len())
352            .finish()
353    }
354}
355
356pub struct Transcription<T, M: Mode, L = ()> {
357    config: M::Config,
358    transcriber: T,
359    layers: L,
360}
361
362pub struct Operation<'a, T, L = ()> {
363    transcriber: &'a T,
364    language: &'a Language,
365    cancel: CancelToken,
366    strategy: Strategy,
367    transcript: Option<Transcript>,
368    layers: &'a L,
369    output: Option<layer::Output>,
370}
371
372use sealed::{Strategy, WholeRecordingInput};
373
374struct ChunkStep {
375    update: Option<TranscriptUpdate>,
376    transcript: Option<Transcript>,
377}
378
379/// Public items here stay unnameable outside this module.
380mod sealed {
381    use super::{CancelToken, Language, Transcriber, TranscriptionError, agreement, native};
382
383    pub enum Strategy {
384        WholeRecording(WholeRecordingInput),
385        Agreement(Box<agreement::AgreementState>),
386        Native(Box<native::NativeState>),
387    }
388
389    #[derive(Clone, Copy)]
390    pub enum WholeRecordingInput {
391        Samples,
392        #[cfg(not(target_family = "wasm"))]
393        Wav,
394    }
395
396    /// What a mode runs, out of reach of the crates that name it.
397    pub trait Sealed {
398        type Config: Copy + Send + Sync + 'static;
399
400        /// The strategy one operation runs. Takes the transcriber because a
401        /// native stream is opened on it.
402        fn strategy<T: Transcriber>(
403            config: Self::Config,
404            transcriber: &T,
405            language: &Language,
406            cancel: &CancelToken,
407        ) -> Result<Strategy, TranscriptionError>;
408    }
409}
410
411/// The transcription strategy a [`Transcription`] runs, carried in its type.
412/// The set is closed: downstream crates can name a mode, never add one.
413#[doc(hidden)]
414pub trait Mode: sealed::Sealed + 'static {
415    const INCREMENTAL: bool;
416}
417
418/// One decode of the whole recording.
419#[derive(Clone, Copy)]
420pub struct Whole;
421
422impl Mode for Whole {
423    const INCREMENTAL: bool = false;
424}
425
426impl sealed::Sealed for Whole {
427    type Config = WholeRecordingInput;
428
429    fn strategy<T: Transcriber>(
430        config: Self::Config,
431        _transcriber: &T,
432        _language: &Language,
433        _cancel: &CancelToken,
434    ) -> Result<Strategy, TranscriptionError> {
435        Ok(Strategy::WholeRecording(config))
436    }
437}
438
439/// Words as they are spoken: LocalAgreement-2 over a growing buffer, or the
440/// engine's own chunked decode.
441#[derive(Clone, Copy)]
442pub struct Streaming;
443
444impl Mode for Streaming {
445    const INCREMENTAL: bool = true;
446}
447
448impl sealed::Sealed for Streaming {
449    type Config = Stream;
450
451    fn strategy<T: Transcriber>(
452        config: Self::Config,
453        transcriber: &T,
454        language: &Language,
455        cancel: &CancelToken,
456    ) -> Result<Strategy, TranscriptionError> {
457        match config {
458            Stream::Agreement(cadence) => Ok(Strategy::Agreement(Box::new(
459                agreement::AgreementState::new(cadence),
460            ))),
461            Stream::Native => match transcriber.open_stream(language, cancel)? {
462                Some(stream) => Ok(Strategy::Native(Box::new(native::NativeState::new(stream)))),
463                None => Err(TranscriptionError::NoNativeStream),
464            },
465        }
466    }
467}
468
469impl<T> Transcription<T, Whole> {
470    pub fn complete_samples(transcriber: T) -> Self {
471        Self {
472            config: WholeRecordingInput::Samples,
473            transcriber,
474            layers: (),
475        }
476    }
477
478    #[cfg(not(target_family = "wasm"))]
479    pub fn complete_wav(transcriber: T) -> Self {
480        Self {
481            config: WholeRecordingInput::Wav,
482            transcriber,
483            layers: (),
484        }
485    }
486
487    /// One decode of the whole recording, handed over as `input` — an
488    /// engine's [`Engine::INPUT`].
489    #[cfg(not(target_family = "wasm"))]
490    pub fn complete(input: AudioInputKind, transcriber: T) -> Self {
491        match input {
492            AudioInputKind::Samples => Self::complete_samples(transcriber),
493            AudioInputKind::Wav => Self::complete_wav(transcriber),
494        }
495    }
496}
497
498impl<T> Transcription<T, Streaming> {
499    /// The escape hatch: agreement over an engine that has not verified a
500    /// cadence, for experiments. A supported streaming path goes through
501    /// [`Transcription::streaming`].
502    pub fn agreement(cadence: Cadence, transcriber: T) -> Self {
503        Self::stream(Stream::Agreement(cadence), transcriber)
504    }
505
506    /// Streams the way `stream` says, over any transcriber. A wrapper around
507    /// an engine passes that engine's [`StreamingEngine::STREAM`] here.
508    pub fn stream(stream: Stream, transcriber: T) -> Self {
509        Self {
510            config: stream,
511            transcriber,
512            layers: (),
513        }
514    }
515}
516
517impl<T: StreamingEngine> Transcription<T, Streaming> {
518    pub fn streaming(engine: T) -> Self {
519        Self::stream(T::STREAM, engine)
520    }
521}
522
523impl<T, M: Mode, L: Layer> Transcription<T, M, L> {
524    pub fn with<N: Layer>(self, layer: N) -> Transcription<T, M, Layered<N, L>> {
525        Transcription {
526            config: self.config,
527            transcriber: self.transcriber,
528            layers: Layered::new(self.layers, layer),
529        }
530    }
531}
532
533impl<T: Transcriber, M: Mode, L: Layer> Transcription<T, M, L> {
534    /// Starts an operation without a cancellation token.
535    pub fn start<'a>(
536        &'a self,
537        language: &'a Language,
538    ) -> Result<Operation<'a, T, L>, TranscriptionError> {
539        self.start_with(language, &CancelToken::never())
540    }
541
542    pub fn start_with<'a>(
543        &'a self,
544        language: &'a Language,
545        cancel: &CancelToken,
546    ) -> Result<Operation<'a, T, L>, TranscriptionError> {
547        check_cancelled(cancel)?;
548        self.transcriber.validate_language(language, cancel)?;
549        check_cancelled(cancel)?;
550        let strategy = M::strategy(self.config, &self.transcriber, language, cancel)?;
551        Ok(Operation {
552            transcriber: &self.transcriber,
553            language,
554            cancel: cancel.clone(),
555            strategy,
556            transcript: None,
557            layers: &self.layers,
558            output: self.layers.enabled().then(layer::Output::default),
559        })
560    }
561
562    pub fn transcribe_with(
563        &self,
564        recording: &mut RecordingReader,
565        language: &Language,
566        cancel: &CancelToken,
567    ) -> Result<Transcript, TranscriptionError> {
568        self.start_with(language, cancel)?.finish(recording)
569    }
570
571    pub fn transcribe(
572        &self,
573        recording: &mut RecordingReader,
574        language: &Language,
575    ) -> Result<Transcript, TranscriptionError> {
576        self.transcribe_with(recording, language, &CancelToken::never())
577    }
578
579    #[cfg(not(target_family = "wasm"))]
580    pub fn transcribe_file(
581        &self,
582        path: impl AsRef<std::path::Path>,
583    ) -> Result<Transcript, TranscriptionError> {
584        let mut recording =
585            RecordingReader::from_wav_file(path).map_err(TranscriptionError::Capture)?;
586        self.transcribe(&mut recording, &Language::Automatic)
587    }
588}
589
590impl<T: Transcriber, L: Layer> Operation<'_, T, L> {
591    pub fn next_finalized(
592        &mut self,
593        recording: &mut RecordingReader,
594    ) -> Result<Option<TranscriptSegment>, TranscriptionError> {
595        loop {
596            match self.next(recording, false)? {
597                Some(TranscriptUpdate::Finalized(segment)) => return Ok(Some(segment)),
598                Some(TranscriptUpdate::Provisional(_)) => {}
599                None => return Ok(None),
600            }
601        }
602    }
603
604    /// The completed transcript, available from the step that ends the
605    /// recording — before the caller has drained the operation. A streaming
606    /// caller checks it so a delivery that fails afterwards cannot lose the
607    /// utterance.
608    pub fn completed(&self) -> Option<&Transcript> {
609        self.transcript.as_ref()
610    }
611
612    /// The earliest stream position this operation may still read. A caller
613    /// done with everything behind it can release that audio.
614    ///
615    /// Whole-recording transcription holds position zero: it hands the
616    /// transcriber the entire recording at once.
617    pub fn retained_from(&self) -> u64 {
618        let strategy = self.strategy_retained_from();
619        self.output
620            .as_ref()
621            .and_then(|output| output.retained_from)
622            .map_or(strategy, |retained| retained.min(strategy))
623    }
624
625    fn strategy_retained_from(&self) -> u64 {
626        match &self.strategy {
627            Strategy::WholeRecording(_) => 0,
628            Strategy::Agreement(state) => state.retained_from(),
629            Strategy::Native(state) => state.retained_from(),
630        }
631    }
632
633    pub fn next_update(
634        &mut self,
635        recording: &mut RecordingReader,
636    ) -> Result<Option<TranscriptUpdate>, TranscriptionError> {
637        self.next(recording, true)
638    }
639
640    /// The next update, with provisional words only when `previews` asks
641    /// for them. Without previews, none are built.
642    pub fn next(
643        &mut self,
644        recording: &mut RecordingReader,
645        previews: bool,
646    ) -> Result<Option<TranscriptUpdate>, TranscriptionError> {
647        let Some(mut output) = self.output.take() else {
648            return self.next_raw_step(recording, previews);
649        };
650        let result = check_cancelled(&self.cancel)
651            .and_then(|()| self.next_processed(recording, previews, &mut output));
652        if result.is_err() {
653            output.finished = true;
654            output.pending.clear();
655            output.text.clear();
656            self.transcript = None;
657        }
658        self.output = Some(output);
659        result
660    }
661
662    fn next_processed(
663        &mut self,
664        recording: &mut RecordingReader,
665        previews: bool,
666        output: &mut layer::Output,
667    ) -> Result<Option<TranscriptUpdate>, TranscriptionError> {
668        loop {
669            if let Some(update) = output.pending.pop_front() {
670                return Ok(Some(update));
671            }
672            if output.finished {
673                return Ok(None);
674            }
675            let update = self.next_raw_step(recording, previews)?;
676            let finished = update.is_none() || self.transcript.is_some();
677            if let Some(update) = update {
678                output.push(update, self.layers, &self.cancel)?;
679            }
680            if finished {
681                // Completion includes queued segments; process that tail before exposing it.
682                output.retained_from = Some(self.strategy_retained_from());
683                while let Some(update) = self.next_raw_step(recording, previews)? {
684                    output.push(update, self.layers, &self.cancel)?;
685                }
686                self.transcript = if output.text.is_empty() {
687                    None
688                } else {
689                    Some(Transcript(std::mem::take(&mut output.text)))
690                };
691                output.finished = true;
692            }
693        }
694    }
695
696    fn next_raw_step(
697        &mut self,
698        recording: &mut RecordingReader,
699        previews: bool,
700    ) -> Result<Option<TranscriptUpdate>, TranscriptionError> {
701        let step = match &mut self.strategy {
702            Strategy::WholeRecording(input) => {
703                if self.transcript.is_some() {
704                    return Ok(None);
705                }
706                let completed = transcribe_complete(
707                    *input,
708                    recording,
709                    self.transcriber,
710                    self.language,
711                    &self.cancel,
712                )?;
713                let segment = TranscriptSegment {
714                    text: completed.as_str().to_owned(),
715                    span: None,
716                };
717                self.transcript = Some(completed);
718                ChunkStep {
719                    update: Some(TranscriptUpdate::Finalized(segment)),
720                    transcript: None,
721                }
722            }
723            Strategy::Agreement(state) => {
724                let cancel = self.cancel.clone();
725                match state.next_with(recording, &cancel, previews, |samples| {
726                    self.transcriber.transcribe_with(
727                        TranscriptionAudio::Samples(samples),
728                        self.language,
729                        &cancel,
730                    )
731                })? {
732                    Some(step) => step,
733                    None => return Ok(None),
734                }
735            }
736            Strategy::Native(state) => match state.next_finalized_with(recording, &self.cancel)? {
737                Some(step) => step,
738                None => return Ok(None),
739            },
740        };
741        if let Some(transcript) = step.transcript {
742            self.transcript = Some(transcript);
743        }
744        Ok(step.update)
745    }
746
747    pub fn finish(
748        &mut self,
749        recording: &mut RecordingReader,
750    ) -> Result<Transcript, TranscriptionError> {
751        while self.next_finalized(recording)?.is_some() {}
752        self.transcript.take().ok_or(TranscriptionError::NoSpeech)
753    }
754}
755
756fn transcribe_complete<T: Transcriber>(
757    input: WholeRecordingInput,
758    recording: &mut RecordingReader,
759    transcriber: &T,
760    language: &Language,
761    cancel: &CancelToken,
762) -> Result<Transcript, TranscriptionError> {
763    check_cancelled(cancel)?;
764    let hypothesis = match input {
765        WholeRecordingInput::Samples => {
766            recording.fill().map_err(TranscriptionError::Capture)?;
767            check_cancelled(cancel)?;
768            transcriber.transcribe_with(
769                TranscriptionAudio::Samples(recording.samples()),
770                language,
771                cancel,
772            )?
773        }
774        #[cfg(not(target_family = "wasm"))]
775        WholeRecordingInput::Wav => {
776            let audio = recording
777                .finished_audio()
778                .map_err(TranscriptionError::Capture)?;
779            check_cancelled(cancel)?;
780            transcriber.transcribe_with(TranscriptionAudio::Wav(audio), language, cancel)?
781        }
782    };
783    check_cancelled(cancel)?;
784    Transcript::new(hypothesis.into_text())
785}
786
787/// [`TranscriptionError::Cancelled`] once `cancel` fires.
788pub fn check_cancelled(cancel: &CancelToken) -> Result<(), TranscriptionError> {
789    if cancel.is_cancelled() {
790        return Err(TranscriptionError::Cancelled);
791    }
792    Ok(())
793}
794
795pub enum TranscriptUpdate {
796    Finalized(TranscriptSegment),
797    Provisional(String),
798}
799
800#[derive(Clone)]
801pub struct TranscriptSegment {
802    text: String,
803    span: Option<std::ops::Range<u64>>,
804}
805
806impl TranscriptSegment {
807    fn spanned(words: &[String], span: Option<std::ops::Range<u64>>) -> Option<Self> {
808        let text = words.join(" ");
809        (!text.is_empty()).then_some(Self { text, span })
810    }
811
812    pub fn as_str(&self) -> &str {
813        &self.text
814    }
815
816    /// Where the segment's audio sits in its recording, when word timestamps
817    /// pin it. Lets a caller measure the silence between two segments.
818    pub fn span(&self) -> Option<std::ops::Range<u64>> {
819        self.span.clone()
820    }
821
822    pub fn into_text(self) -> String {
823        self.text
824    }
825}
826
827impl std::fmt::Debug for TranscriptSegment {
828    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
829        formatter
830            .debug_struct("TranscriptSegment")
831            .field("bytes", &self.text.len())
832            .field("span", &self.span)
833            .finish()
834    }
835}
836
837#[derive(Debug, Clone, PartialEq, Eq)]
838pub enum Language {
839    Automatic,
840    Specific(LanguageCode),
841}
842
843impl Language {
844    const AUTOMATIC: &str = "auto";
845
846    pub fn as_str(&self) -> &str {
847        match self {
848            Self::Automatic => Self::AUTOMATIC,
849            Self::Specific(code) => code.as_str(),
850        }
851    }
852}
853
854/// Reads back what [`Language::as_str`] writes.
855impl std::str::FromStr for Language {
856    type Err = LanguageCodeError;
857
858    fn from_str(value: &str) -> Result<Self, Self::Err> {
859        if value.trim().eq_ignore_ascii_case(Self::AUTOMATIC) {
860            return Ok(Self::Automatic);
861        }
862        LanguageCode::new(value).map(Self::Specific)
863    }
864}
865
866#[derive(Debug, Clone, PartialEq, Eq)]
867pub struct LanguageCode(String);
868
869impl LanguageCode {
870    pub fn english() -> Self {
871        Self("en".to_owned())
872    }
873
874    pub fn new(value: impl Into<String>) -> Result<Self, LanguageCodeError> {
875        let value = value.into();
876        let value = value.trim();
877        if value.is_empty() {
878            return Err(LanguageCodeError::Empty);
879        }
880        if value.len() > MAX_LANGUAGE_CODE_BYTES {
881            return Err(LanguageCodeError::TooLong {
882                actual_bytes: value.len(),
883                maximum_bytes: MAX_LANGUAGE_CODE_BYTES,
884            });
885        }
886        let mut segments = value.split('-');
887        let valid_primary = segments.next().is_some_and(|segment| {
888            !segment.is_empty() && segment.bytes().all(|byte| byte.is_ascii_alphabetic())
889        });
890        let valid_subtags = segments.all(|segment| {
891            !segment.is_empty() && segment.bytes().all(|byte| byte.is_ascii_alphanumeric())
892        });
893        if !valid_primary || !valid_subtags {
894            return Err(LanguageCodeError::Malformed(value.to_owned()));
895        }
896        Ok(Self(value.to_ascii_lowercase()))
897    }
898
899    pub fn as_str(&self) -> &str {
900        &self.0
901    }
902}
903
904#[derive(Debug, thiserror::Error)]
905pub enum LanguageCodeError {
906    #[error("recognition language code must not be empty")]
907    Empty,
908
909    #[error(
910        "recognition language code contains {actual_bytes} bytes; maximum is {maximum_bytes} bytes"
911    )]
912    TooLong {
913        actual_bytes: usize,
914        maximum_bytes: usize,
915    },
916
917    #[error("recognition language code is malformed: {0:?}")]
918    Malformed(String),
919}
920
921#[derive(Clone, PartialEq, Eq)]
922pub struct Transcript(String);
923
924impl Transcript {
925    pub fn new(text: impl Into<String>) -> Result<Self, TranscriptionError> {
926        let text = text.into();
927        let text = text.trim();
928        if text.is_empty() {
929            return Err(TranscriptionError::NoSpeech);
930        }
931        Ok(Self(text.to_owned()))
932    }
933
934    pub fn as_str(&self) -> &str {
935        &self.0
936    }
937
938    pub fn into_text(self) -> String {
939        self.0
940    }
941}
942
943impl std::fmt::Debug for Transcript {
944    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
945        formatter
946            .debug_struct("Transcript")
947            .field("bytes", &self.0.len())
948            .finish()
949    }
950}
951
952#[derive(Debug, thiserror::Error)]
953pub enum TranscriptionError {
954    #[error("the {backend} model cannot be used")]
955    Model {
956        backend: &'static str,
957        #[source]
958        source: ModelError,
959    },
960
961    #[error("failed to load the {backend} model {model}")]
962    Load {
963        backend: &'static str,
964        model: String,
965        #[source]
966        source: BackendError,
967    },
968
969    #[error("the recording being transcribed failed")]
970    Capture(#[source] CaptureError),
971
972    #[error("{backend} inference failed")]
973    Inference {
974        backend: &'static str,
975        #[source]
976        source: BackendError,
977    },
978
979    #[error("{backend} cannot transcribe {language:?}")]
980    UnsupportedLanguage {
981        backend: &'static str,
982        language: String,
983    },
984
985    #[error("{backend} needs a UTF-8 recording path; got {}", path.display())]
986    UnsupportedAudioPath {
987        backend: &'static str,
988        path: PathBuf,
989    },
990
991    #[error("{backend} requires {required}, but received {received}")]
992    UnsupportedAudioInput {
993        backend: &'static str,
994        received: AudioInputKind,
995        required: AudioInputKind,
996    },
997
998    #[error("the {backend} backend is not available on {}", std::env::consts::OS)]
999    UnsupportedBackend { backend: &'static str },
1000
1001    #[error("the {backend} backend has no gpu support on {}", std::env::consts::OS)]
1002    UnsupportedDevice { backend: &'static str },
1003
1004    #[error("the engine was asked to stream natively but offers no stream")]
1005    NoNativeStream,
1006
1007    #[error("the model loader stopped unexpectedly")]
1008    LoadPanicked,
1009
1010    #[error("transcription was cancelled")]
1011    Cancelled,
1012
1013    #[error("no speech was transcribed")]
1014    NoSpeech,
1015
1016    #[error("transcript processing failed")]
1017    Processing(#[source] layer::Error),
1018}
1019
1020/// The shared model validators, answering in this capability's vocabulary.
1021#[cfg(not(target_family = "wasm"))]
1022pub fn validate_model_file(
1023    path: PathBuf,
1024    backend: &'static str,
1025) -> Result<PathBuf, TranscriptionError> {
1026    crate::backend::validate_model_file(path)
1027        .map_err(|source| TranscriptionError::Model { backend, source })
1028}
1029
1030#[cfg(not(target_family = "wasm"))]
1031pub fn validate_model_directory(
1032    path: PathBuf,
1033    backend: &'static str,
1034    required: &[&'static str],
1035) -> Result<PathBuf, TranscriptionError> {
1036    crate::backend::validate_model_directory(path, required)
1037        .map_err(|source| TranscriptionError::Model { backend, source })
1038}
1039
1040// Accept automatic detection and English, regional variants included; reject
1041// other explicit languages.
1042pub fn english_only(backend: &'static str, language: &Language) -> Result<(), TranscriptionError> {
1043    let Language::Specific(code) = language else {
1044        return Ok(());
1045    };
1046    if code.as_str().split('-').next() == Some(LanguageCode::english().as_str()) {
1047        return Ok(());
1048    }
1049    Err(TranscriptionError::UnsupportedLanguage {
1050        backend,
1051        language: code.as_str().to_owned(),
1052    })
1053}
1054
1055#[cfg(test)]
1056mod word_timing_tests {
1057    use super::*;
1058
1059    fn token(text: &str, start: f32, end: f32) -> (&str, TimedWord) {
1060        (text, TimedWord { start, end })
1061    }
1062
1063    #[test]
1064    fn tokens_group_into_words_at_the_sentencepiece_marker() {
1065        let tokens = [
1066            token("\u{2581}hel", 0.0, 0.2),
1067            token("lo", 0.2, 0.4),
1068            token(",", 0.4, 0.5),
1069            token("\u{2581}world", 0.6, 0.9),
1070        ];
1071        let words = word_timings(tokens);
1072        assert_eq!(words.len(), 2);
1073        assert_eq!((words[0].start, words[0].end), (0.0, 0.5));
1074        assert_eq!((words[1].start, words[1].end), (0.6, 0.9));
1075    }
1076
1077    #[test]
1078    fn tokens_group_into_words_at_a_leading_space() {
1079        let tokens = [
1080            token(" hel", 0.0, 0.2),
1081            token("lo", 0.2, 0.4),
1082            token(",", 0.4, 0.5),
1083            token(" world", 0.6, 0.9),
1084        ];
1085        let words = word_timings(tokens);
1086        assert_eq!(words.len(), 2);
1087        assert_eq!((words[0].start, words[0].end), (0.0, 0.5));
1088        assert_eq!((words[1].start, words[1].end), (0.6, 0.9));
1089    }
1090
1091    #[test]
1092    fn word_timings_count_what_the_text_splits_into() {
1093        // A bare marker, a digit run, and a token holding two words.
1094        let tokens = [
1095            token("\u{2581}", 0.0, 0.1),
1096            token("\u{2581}1", 0.1, 0.2),
1097            token("8", 0.2, 0.3),
1098            token("9", 0.3, 0.4),
1099            token("\u{2581}years old", 0.4, 0.8),
1100            token(".", 0.8, 0.9),
1101        ];
1102        let text: String = tokens
1103            .iter()
1104            .map(|(text, _)| text.replace('\u{2581}', " "))
1105            .collect();
1106        let words = word_timings(tokens);
1107        assert_eq!(words.len(), text.split_whitespace().count());
1108        assert_eq!((words[0].start, words[0].end), (0.1, 0.4));
1109        assert_eq!((words[2].start, words[2].end), (0.4, 0.9));
1110    }
1111
1112    #[test]
1113    fn a_word_timing_rounds_to_the_nearest_sample() {
1114        // 0.10003 s is 1600.48 samples and 0.20004 s is 3200.64.
1115        let word = TimedWord {
1116            start: 0.100_03,
1117            end: 0.200_04,
1118        };
1119
1120        assert_eq!(word.span_from(1_000), 2_600..4_201);
1121    }
1122}
1123
1124#[cfg(test)]
1125mod language_tests {
1126    use super::*;
1127
1128    const BACKEND: &str = "test-backend";
1129
1130    #[test]
1131    fn detection_and_english_are_accepted() {
1132        assert!(english_only(BACKEND, &Language::Automatic).is_ok());
1133        assert!(english_only(BACKEND, &Language::Specific(LanguageCode::english())).is_ok());
1134
1135        let regional = LanguageCode::new("en-GB").expect("en-GB is a valid code");
1136        assert!(english_only(BACKEND, &Language::Specific(regional)).is_ok());
1137    }
1138
1139    #[test]
1140    fn language_parses_what_as_str_writes() {
1141        for language in [
1142            Language::Automatic,
1143            Language::Specific(LanguageCode::new("en-GB").expect("en-GB is a valid code")),
1144        ] {
1145            assert_eq!(language.as_str().parse::<Language>().ok(), Some(language));
1146        }
1147        assert_eq!(" AUTO".parse::<Language>().ok(), Some(Language::Automatic));
1148        assert!(matches!(
1149            "en_GB".parse::<Language>(),
1150            Err(LanguageCodeError::Malformed(_))
1151        ));
1152    }
1153
1154    #[test]
1155    fn another_language_is_refused_by_name() {
1156        let code = LanguageCode::new("de").expect("de is a valid code");
1157
1158        let error = english_only(BACKEND, &Language::Specific(code))
1159            .expect_err("another language must be refused");
1160
1161        assert!(matches!(
1162            error,
1163            TranscriptionError::UnsupportedLanguage { language, .. } if language == "de"
1164        ));
1165    }
1166}
1167
1168#[cfg(test)]
1169mod coordinator_tests {
1170    use std::sync::{Arc, Mutex};
1171
1172    use crate::audio::testing::recorded;
1173
1174    use super::*;
1175
1176    struct FakeTranscriber {
1177        response: &'static str,
1178        reject_language: bool,
1179        inputs: Mutex<Vec<AudioInputKind>>,
1180    }
1181
1182    impl FakeTranscriber {
1183        fn answering(response: &'static str) -> Self {
1184            Self {
1185                response,
1186                reject_language: false,
1187                inputs: Mutex::new(Vec::new()),
1188            }
1189        }
1190
1191        fn inputs(&self) -> Vec<AudioInputKind> {
1192            self.inputs
1193                .lock()
1194                .unwrap_or_else(|poisoned| poisoned.into_inner())
1195                .clone()
1196        }
1197    }
1198
1199    impl Transcriber for FakeTranscriber {
1200        fn validate_language(
1201            &self,
1202            language: &Language,
1203            _cancel: &CancelToken,
1204        ) -> Result<(), TranscriptionError> {
1205            if self.reject_language {
1206                return Err(TranscriptionError::UnsupportedLanguage {
1207                    backend: "fake",
1208                    language: language.as_str().to_owned(),
1209                });
1210            }
1211            Ok(())
1212        }
1213
1214        fn transcribe_with(
1215            &self,
1216            audio: TranscriptionAudio<'_>,
1217            _language: &Language,
1218            _cancel: &CancelToken,
1219        ) -> Result<Hypothesis, TranscriptionError> {
1220            self.inputs
1221                .lock()
1222                .unwrap_or_else(|poisoned| poisoned.into_inner())
1223                .push(audio.kind());
1224            Ok(Hypothesis::new(self.response))
1225        }
1226    }
1227
1228    /// An engine that streams natively: every chunk answers "hello".
1229    struct Chunked;
1230
1231    struct HelloStream;
1232
1233    impl ChunkStream for HelloStream {
1234        fn chunk_samples(&self) -> usize {
1235            2
1236        }
1237
1238        fn feed(&mut self, chunk: &[f32]) -> Result<Emitted, TranscriptionError> {
1239            Ok(Emitted {
1240                words: Hypothesis::new(match chunk.iter().any(|sample| *sample != 0.0) {
1241                    true => "hello",
1242                    false => "",
1243                }),
1244                joins_previous: false,
1245            })
1246        }
1247    }
1248
1249    impl Transcriber for Chunked {
1250        fn validate_language(
1251            &self,
1252            _language: &Language,
1253            _cancel: &CancelToken,
1254        ) -> Result<(), TranscriptionError> {
1255            Ok(())
1256        }
1257
1258        fn transcribe_with(
1259            &self,
1260            _audio: TranscriptionAudio<'_>,
1261            _language: &Language,
1262            _cancel: &CancelToken,
1263        ) -> Result<Hypothesis, TranscriptionError> {
1264            unreachable!("a native stream never decodes a bounded input")
1265        }
1266
1267        fn open_stream(
1268            &self,
1269            _language: &Language,
1270            _cancel: &CancelToken,
1271        ) -> Result<Option<Box<dyn ChunkStream>>, TranscriptionError> {
1272            Ok(Some(Box::new(HelloStream)))
1273        }
1274    }
1275
1276    #[test]
1277    fn a_boxed_or_shared_transcriber_still_opens_its_stream() {
1278        let (_directory, mut boxed_recording) = recorded(&[1, 2, 3, 4]);
1279        let (_directory, mut shared_recording) = recorded(&[1, 2, 3, 4]);
1280        let boxed: Box<dyn Transcriber> = Box::new(Chunked);
1281        let shared = Arc::new(Chunked);
1282
1283        let from_boxed = Transcription::stream(Stream::Native, boxed)
1284            .transcribe(&mut boxed_recording, &Language::Automatic)
1285            .expect("the boxed stream opens");
1286        let from_shared = Transcription::stream(Stream::Native, shared)
1287            .transcribe(&mut shared_recording, &Language::Automatic)
1288            .expect("the shared stream opens");
1289
1290        assert_eq!(from_boxed.as_str(), "hello hello");
1291        assert_eq!(from_shared.as_str(), "hello hello");
1292    }
1293
1294    #[test]
1295    fn completion_is_available_before_pending_final_segments_and_preview_clear() {
1296        struct Ending(std::sync::atomic::AtomicBool);
1297        impl Transcriber for Ending {
1298            fn validate_language(
1299                &self,
1300                _: &Language,
1301                _: &CancelToken,
1302            ) -> Result<(), TranscriptionError> {
1303                Ok(())
1304            }
1305            fn transcribe_with(
1306                &self,
1307                _: TranscriptionAudio<'_>,
1308                _: &Language,
1309                _: &CancelToken,
1310            ) -> Result<Hypothesis, TranscriptionError> {
1311                Ok(Hypothesis::new(
1312                    if self.0.swap(true, std::sync::atomic::Ordering::Relaxed) {
1313                        "one two three"
1314                    } else {
1315                        "one two"
1316                    },
1317                ))
1318            }
1319        }
1320        let (mut recording, mut capture) = crate::audio::Recording::open();
1321        let transcription = Transcription::agreement(
1322            Cadence::from_millis(100, 20_000),
1323            Ending(std::sync::atomic::AtomicBool::new(false)),
1324        );
1325        let mut operation = transcription.start(&Language::Automatic).expect("starts");
1326        capture.push(&[0.5; 1600]);
1327        assert!(
1328            matches!(operation.next_update(&mut recording).expect("preview"), Some(TranscriptUpdate::Provisional(text)) if text == "one two")
1329        );
1330        capture.push(&[0.5; 1600]);
1331        capture.finish();
1332        recording.fill().expect("capture ended");
1333        assert!(
1334            matches!(operation.next_update(&mut recording).expect("final"), Some(TranscriptUpdate::Finalized(segment)) if segment.as_str() == "one two")
1335        );
1336        assert_eq!(
1337            operation.completed().expect("already complete").as_str(),
1338            "one two three"
1339        );
1340        assert!(
1341            matches!(operation.next_update(&mut recording).expect("tail"), Some(TranscriptUpdate::Finalized(segment)) if segment.as_str() == "three")
1342        );
1343        assert!(
1344            matches!(operation.next_update(&mut recording).expect("clear"), Some(TranscriptUpdate::Provisional(text)) if text.is_empty())
1345        );
1346        assert!(
1347            operation
1348                .next_update(&mut recording)
1349                .expect("ended")
1350                .is_none()
1351        );
1352    }
1353
1354    #[test]
1355    fn native_and_whole_updates_are_finalized_only() {
1356        fn check<T: Transcriber, M: Mode>(transcription: Transcription<T, M>, expected: &str) {
1357            let (_directory, mut recording) = recorded(&[1, 2, 3, 4]);
1358            let mut operation = transcription.start(&Language::Automatic).expect("starts");
1359            let mut segments = Vec::new();
1360            while let Some(update) = operation.next_update(&mut recording).expect("reads") {
1361                let TranscriptUpdate::Finalized(segment) = update else {
1362                    panic!("only agreement exposes provisional words");
1363                };
1364                segments.push(segment.into_text());
1365            }
1366            assert_eq!(segments.join(" "), expected);
1367            assert_eq!(operation.completed().expect("completed").as_str(), expected);
1368            assert_eq!(
1369                operation.finish(&mut recording).expect("finished").as_str(),
1370                expected
1371            );
1372        }
1373        check(
1374            Transcription::stream(Stream::Native, Chunked),
1375            "hello hello",
1376        );
1377        check(
1378            Transcription::complete_samples(FakeTranscriber::answering("hello there")),
1379            "hello there",
1380        );
1381    }
1382
1383    #[test]
1384    fn streaming_natively_without_a_stream_is_an_error() {
1385        let transcription =
1386            Transcription::stream(Stream::Native, FakeTranscriber::answering("hello"));
1387
1388        let outcome = transcription.start(&Language::Automatic).map(|_| ());
1389
1390        assert!(matches!(outcome, Err(TranscriptionError::NoNativeStream)));
1391    }
1392
1393    #[test]
1394    fn complete_samples_invokes_the_engine_once() {
1395        let (_directory, mut recording) = recorded(&[1, 2, 3]);
1396        let transcriber = Arc::new(FakeTranscriber::answering("hello there"));
1397        let transcription = Transcription::complete_samples(Arc::clone(&transcriber));
1398
1399        let transcript = transcription
1400            .transcribe(&mut recording, &Language::Automatic)
1401            .expect("the fake returns speech");
1402
1403        assert_eq!(transcript.as_str(), "hello there");
1404        assert_eq!(transcriber.inputs(), [AudioInputKind::Samples]);
1405    }
1406
1407    #[cfg(not(target_family = "wasm"))]
1408    #[test]
1409    fn complete_wav_preserves_the_file_input_boundary() {
1410        let (_directory, mut recording) = recorded(&[1, 2, 3]);
1411        let transcriber = Arc::new(FakeTranscriber::answering("hello there"));
1412        let transcription = Transcription::complete_wav(Arc::clone(&transcriber));
1413
1414        transcription
1415            .transcribe(&mut recording, &Language::Automatic)
1416            .expect("the fake returns speech");
1417
1418        assert_eq!(transcriber.inputs(), [AudioInputKind::Wav]);
1419        assert!(recording.samples().is_empty());
1420    }
1421
1422    #[test]
1423    fn an_empty_hypothesis_becomes_terminal_no_speech() {
1424        let (_directory, mut recording) = recorded(&[1, 2, 3]);
1425        let transcription = Transcription::complete_samples(FakeTranscriber::answering("   "));
1426
1427        let outcome = transcription.transcribe(&mut recording, &Language::Automatic);
1428
1429        assert!(matches!(outcome, Err(TranscriptionError::NoSpeech)));
1430    }
1431
1432    #[test]
1433    fn cancellation_prevents_complete_inference() {
1434        let (_directory, mut recording) = recorded(&[1, 2, 3]);
1435        let transcriber = Arc::new(FakeTranscriber::answering("unreachable"));
1436        let transcription = Transcription::complete_samples(Arc::clone(&transcriber));
1437        let cancel = CancelToken::new();
1438        cancel.cancel();
1439
1440        let outcome = transcription.transcribe_with(&mut recording, &Language::Automatic, &cancel);
1441
1442        assert!(matches!(outcome, Err(TranscriptionError::Cancelled)));
1443        assert!(transcriber.inputs().is_empty());
1444    }
1445
1446    #[test]
1447    fn language_rejection_prevents_complete_inference() {
1448        let (_directory, mut recording) = recorded(&[1, 2, 3]);
1449        let mut fake = FakeTranscriber::answering("unreachable");
1450        fake.reject_language = true;
1451        let transcriber = Arc::new(fake);
1452        let transcription = Transcription::complete_samples(Arc::clone(&transcriber));
1453
1454        let outcome = transcription.transcribe(&mut recording, &Language::Automatic);
1455
1456        assert!(matches!(
1457            outcome,
1458            Err(TranscriptionError::UnsupportedLanguage { .. })
1459        ));
1460        assert!(transcriber.inputs().is_empty());
1461    }
1462
1463    #[test]
1464    fn sensitive_transcription_types_redact_their_text() {
1465        let secret = "words that must not reach diagnostics";
1466        let hypothesis = Hypothesis::new(secret);
1467        let transcript = Transcript::new(secret).expect("the text is not empty");
1468        let segment =
1469            TranscriptSegment::spanned(&[secret.to_owned()], None).expect("the text is not empty");
1470
1471        assert!(!format!("{hypothesis:?}").contains(secret));
1472        assert!(!format!("{transcript:?}").contains(secret));
1473        assert!(!format!("{segment:?}").contains(secret));
1474    }
1475}
aria_kit/transcription/mod.rs

aria_kit/transcription/
mod.rs
