Standard speech recognition often fails in environments like noisy airports not by mishearing, but by faithfully transcribing crisp background announcements into personal documents. While commercial tools like Krisp isolate audio beforehand, target-speaker automatic speech recognition integrates speaker identity directly into the model via thirty-to-sixty-second enrollments. Incorporating learned speaker embeddings or frame-level state labels allows models to steer recognition toward an authorized user. Crucially, research from NTT and Brno reveals that a superior speaker-verification vector does not always guarantee better target recognition, as optimal conditioning vectors shift depending on competing background voices. Additionally, systems must be explicitly trained on target-absent stretches, teaching the recognizer to return entirely empty transcripts when an enrolled speaker stops talking.
Episode 6297 manual-script
Teaching Dictation Models to Ignore Other Voices
Evaluating Whisper on non-speech clips revealed entire invented phrases in roughly one percent of transcriptions.
22:03 Transcript
Episode Details
- Episode ID
- MWP-6499
- Published
- Duration
- 22:03
- Audio
- Direct link
- Pipeline
- V5.22
- TTS Engine
-
chatterbox-regular - Script Writing Agent
- GPT-6.1 Sol
AI-Generated Content: This podcast is created using AI personas. Please verify any important information independently.
The prompt
The prompt that started this episode, as sent in (production notes for the hosts removed).
Could a voice app enrol the user with 30-60 seconds of speech, compute a speaker embedding, and use it inside the ASR model (via attention or conditioning) so only the user's speech is transcribed and a clear PA announcement like 'Delta flight 10 to Newark' produces nothing?
Mentions
- Bag of Hallucinations Catalog of recurring Whisper non-speech outputs
- DiCoP Parakeet-based cousin of DiCoW
- DiCoW Frame-label conditioned target-speaker ASR model
- Krisp AI noise suppression software
- Libri2Mix Speech separation benchmark dataset
- Parakeet ASR model for dictation transcription
- Speaker-querying Whisper Enrollment-conditioned Whisper variant for target speaker
- Target-speaker unlearning Inverse task of excluding a speaker from transcription
- Whisper OpenAI's speech-to-text model
- WSJ mixtures Speech separation benchmark from Wall Street Journal corpus
Downloads
Transcript (TXT)
Plain text transcript file
Episode Book (PDF)
The episode's record — date, duration, models, sources — with the full transcript
You Might Also Like
Never miss an episode
New episodes drop daily — subscribe on your favorite platform
New to the show? Start here#6297: Teaching Dictation Models to Ignore Other Voices
Better noise cancellation is the standard answer to dictation going wrong in an airport. Daniel's example turns that answer inside out. The offending sound is a perfectly clear announcement, and he wants the speech model itself to know it belongs to somebody else.
His proposal is to record thirty to sixty seconds of your voice, turn that into a speaker embedding, and use it inside recognition, perhaps through attention, so only your speech becomes text. Ten seconds of Daniel dictating should appear. The announcement that follows, Ladies and gentlemen, we are now ready to board Delta flight ten to Newark, should produce nothing.
The airport has delivered an unusually legible input. Recognition succeeds, and the application fails.
Newark has entered the document uninvited.
Target-speaker ASR is the name for the architecture Daniel is imagining. The answer is yes, you can build speaker selection into recognition. The harder answer concerns what information you give it, where that information enters, and what happens when the intended speaker stops talking.
He's interested mainly in one authorised user. A family meeting with another authorised voice is an extension, not the main event. I want to hold you to that, Herman. Your meeting diagram already has chairs.
Two chairs. I've shown restraint.
A speech activity detector will happily open for that announcement. A quality score may prefer it to Daniel's dictation. Loudness might favour it too. None of those signals tells the application whose words belong in the document.
Your instinct would be to improve the microphone placement.
My instinct would be to remove the airport. Placement still helps, but you're right about the limit. Once clean speech reaches the recogniser, the system needs a target definition. Otherwise, faithfully writing it down is a reasonable response to the task it was given.
I suspect people imagine the app already knows the microphone is theirs, therefore the words must be theirs. Ownership of the laptop has somehow become an acoustic feature.
The microphone doesn't get the purchase receipt. Target-speaker systems add another input that identifies whom to follow. Most published work concentrates on simultaneous speakers, using mixtures such as Libri2Mix or recordings of meetings. Separating two overlapping voices is difficult, so that's where much of the effort goes.
Daniel's sequence has a different shape. He speaks, stops, and the announcement starts. There's no simultaneous voice to separate during the announcement. The desired output for a whole stretch of excellent speech is empty.
Empty is a surprisingly ambitious feature.
Commercial voice isolation already exists. Krisp, for one, sells isolation that sits ahead of speech recognition. The usual commercial arrangement cleans or selects the audio before the recogniser receives it.
Which answers whether someone can sell the experience, but leaves Daniel's architectural question open. He wants the recogniser to carry the intention, rather than receive audio somebody else has approved.
Would you care where the selection happens, Corn, if the experience were identical?
As a user, barely. As someone debugging the failure, enormously. If the airport survives the filter, can the recogniser still refuse it? I'd rather have that decision represented somewhere than hope the unwanted voice has vanished.
That distinction matters. Integration gives recognition a chance to learn which acoustic evidence deserves words, without requiring an intermediate waveform that sounds clean to us.
A conventional recogniser learns a mapping from speech audio to text. Whisper-style foundation models weren't generally trained with an enrolled user and an instruction to reject every other voice. Their architecture doesn't come with a personal ownership test.
Training it on more airport announcements could make this worse.
It could make the announcements easier to transcribe. You need the objective to change. The same recording might have Daniel's transcript as the correct answer when Daniel is selected, and the announcer's transcript when the announcer is selected.
Same sound, different correct document. That makes the enrollment input part of the task definition, rather than a little accuracy enhancement.
Yes... and we should separate this from hallucination. Whisper can produce phrases when there was no speech corresponding to them. Its internal representations also retain information about non-speech sounds. Noise robustness doesn't mean noise disappears inside the model.
One Polish team ran over three hundred thousand non-speech clips through it and collected recurring outputs into a Bag of Hallucinations. Separate evaluations found whole invented phrases in roughly one percent of transcriptions. Those findings concern particular evaluations, not a universal rate for every recording.
The Delta announcement needs no invention. It's an accurate transcription of an unwanted event.
Right. Well... accurate if the words are recognised correctly. Even perfect accuracy on that voice would violate Daniel's request. Post-processing for suspicious invented phrases won't remove a legitimate boarding announcement reliably.
Especially if I'm actually dictating a boarding announcement. The words themselves cannot be the blacklist.
Enrollment provides acoustic identity instead. A speaker encoder processes the sample and compresses it into a fixed-size vector. Training encourages samples from the same speaker to land near each other despite different sentences, with different speakers farther apart.
It has to forget the sentence while remembering the speaker.
Broadly. It won't forget every trace of content or recording conditions. Those are some of the complications. A comparison of target-speaker systems used vectors with five hundred and twelve dimensions. They're learned coordinates, not a row of labelled measurements for accent, pitch and throat size.
No little box marked donkey.
The absence has not impaired my representation.
Thirty seconds sounds short for a permanent description of a person. Though I suppose your voice gives them plenty to work with.
Mine is generously documented. Thirty to sixty seconds of reasonably clean, varied speech is plenty for a useful enrollment. More precisely, getting a stable speaker representation from that amount is plausible and routine. Getting flawless rejection from it is a separate demand.
Here's the part I think you'll object to before I finish. A better speaker-verification embedding doesn't necessarily produce better target-speaker recognition.
You were going to sell me the vector before revealing that.
I was going to show you the elegant mechanism first.
There it is.
The NTT and Brno comparison tested speaker representations across recognition, extraction and personal speech detection. Strong verification performance didn't reliably predict strong performance on those tasks. An ideal identity code could outperform enrollment-derived vectors.
An identity code meaning the experiment already knows which person it is? Not a miraculous microphone that assigns everybody a flawless number.
Correct. It's a controlled reference. It helps distinguish trouble in representing identity from trouble using identity. More awkwardly, when they optimised the conditioning vector for the target task, its best position changed with the competing speech.
I remain the same person, but the useful description of me changes depending on who's interrupting.
The vector needs to emphasise whatever distinguishes you from that particular interference. A fixed enrollment remains useful. The result challenges the assumption that one universal voiceprint is automatically the best steering signal forever.
I buy that, with a reservation. Daniel doesn't need the mathematically best vector in every room. He needs one good enough to reject the airport without demanding a fresh audition beside every gate.
That's fair... useful and optimal are different thresholds. Would you prefer the system to adapt its conditioning during a session?
Only if it keeps the original identity anchored. Otherwise it might adapt beautifully to the announcer while I'm buying coffee. Improvement needs a definition of whom it's improving for.
That's why session context and target assignment have to remain distinct. The alternative approach gives the recogniser frame-level labels: silence, target speech, non-target speech, or overlap.
Hilbert's Anteater Trivia. Emas National Park, in central Brazil, is known for its giant anteaters. I have never been, which puts me ahead of most park rangers.
A convincing case for staying home.
Those labels let the recognition layers respond differently to each state. We went through DiCoW and its Parakeet cousin, DiCoP, last time: small learned transformations per label, and they work because another component has already made the difficult judgement of who is speaking when. No speaker embedding in sight.
The distinction is attractive, though. The recogniser gets told what role this sound has now. It doesn't have to discover an entire identity catalogue before deciding whether a word belongs.
Embedding-conditioned models take a different route. Speaker-querying Whisper uses learned queries and enrollment information to draw speaker-specific features from the mixture. Those features steer recognition towards the selected voice.
Attention gets evidence of whom to follow, rather than a command written in English saying ignore everybody else.
Training makes that relationship meaningful. Just attaching a vector to an existing decoder doesn't give the coordinates a shared meaning. You need examples, a way to inject the information, and an objective that rewards the selected transcript.
Otherwise it's five hundred and twelve numbers wearing a visitor badge.
You could inject the conditioning alongside acoustic features, through transformations inside the encoder, or through attention that consults speaker information. Daniel's decoding-stage intuition is sound in spirit, but the useful intervention often happens before final token selection.
Waiting until the model has decided the sentence is about Delta seems late. I would want identity attached to the acoustic evidence before the language model gets enthusiastic about finishing the destination.
Earlier access helps preserve the association between words and voice. The canonical design has an auxiliary enrollment network and a recognising network taking both its representation and the live audio. It can be trained end to end, or adapted from an existing recogniser.
The speaker-querying approach gets about fifteen percent word error on Libri2Mix and around four percent on WSJ mixtures. Different benchmarks, different difficulty. Those results establish that enrollment can steer transcription, not that an airport app has been solved.
What does the training example look like when Daniel is silent? I think that matters more than the clever attention layer.
It should include non-target-only stretches whose desired transcript is empty. Also target-only stretches, overlap, pauses and transitions. If training overwhelmingly says there are two voices and one must be selected, it may not adequately teach the case where the target is absent.
The model must be allowed to find nobody.
Yes. No... more than allowed. It must be rewarded for that decision in the right circumstances. Daniel's PA case may be easier for a segment-level identity gate because nothing overlaps. Yet it tests target absence, something an overlap benchmark score can conceal.
I'd score the airport intrusion separately. A transcript with every word of my dictation intact and one unwanted sentence could look respectable on a broad accuracy average. I'd still delete the app.
What would you put beside word error, then?
How much non-target speech reaches the document, and how much of mine disappears. Separately. Otherwise a very quiet system wins by refusing to transcribe anyone, which is a remarkably expensive notepad.
Streaming adds the timing dimension. An offline system can inspect later evidence before deciding whose voice it heard. A live dictation system must either wait, emit provisional text, or risk committing the wrong speaker's opening words.
Dimmed text doesn't feel provisional when it says I'm boarding.
A streaming target-speaker transducer with dual attention remained around sixteen percent word error under contaminated enrollment, while comparison approaches exceeded seventy-five. That demonstrates robustness in its tested conditions. It doesn't give us end-to-end delay for a dictation product.
You can call a system streaming while leaving me unsure how long the first acceptable word takes.
Processing incrementally and feeling immediate are different properties. I'm not sure what delay would be acceptable here. Would you tolerate a short hold before each new speaker?
Before a speaker change, perhaps. Before every phrase after a breath, less so. I pause to think, Herman. I don't want my own software treating contemplation as a suspicious new identity.
It would have ample test material.
You may use the gaps to recover from your explanations.
Another complication is that enrollment isn't always clean. Self-enrolled systems can locate a passage in the conversation where the selected speaker is most active, then use that passage through cross-attention in the encoder. One extension roughly halved its speaker-attributed error measure against its predecessor on its benchmark.
Useful when nobody remembered to record the ceremonial voice sample. Though somebody still needs to decide which participant is the user.
Self-enrollment finds representative audio. It doesn't read intention. Additional cues can help: in one dual-cue evaluation, enrollment alone gave roughly twenty-nine percent character error, while enrollment plus a short text cue brought it below nine percent.
Text cues sound useful when you know what someone might say. My next dictated sentence is often news to me.
They're task-dependent. The result says voice enrollment needn't be the strongest available cue. It doesn't mean dictation should require a script.
The family-meeting exception stays small, then. Add another authorised identity, permit either, and perhaps preserve who said what if the document needs it.
Diarization tracks who spoke when across participants. Daniel's ordinary case is target membership, an authorised-user decision. With multiple authorised voices, diarization and attribution become more useful. A product would likely split it the same way: a voice profile for personal dictation and a separate meeting mode.
I still favour the boring upstream component for shipping this soon. I know you're preparing to accuse me of resisting elegance.
I was preparing to agree. A pre-filter works with existing recognisers. Rebuilding their conditioning path means model access, training, evaluation and maintenance every time the base model changes.
You've installed enough fresh models to qualify as a maintenance event.
The installation is often the pleasant part. Krisp's Voice Isolation sits ahead of transcription and reports forty-six percent lower word error with fifteen milliseconds of processing latency on a CPU. Those are company evaluation claims, and the latency belongs to isolation, not the entire app.
Its tests covered ten recognition engines and over sixteen hundred real-world recordings. That establishes a practical path without proving anything about Daniel's in-model proposal.
They also inherit a harsh limit. If the filter removes my quiet word, the recogniser never gets a vote.
An integrated model could use phonetic and speaker evidence jointly instead. Conversely, a separately inspectable filter makes debugging easier. Neither arrangement automatically wins on clean-speech preservation.
Have we actually got the exact airport test evaluated anywhere?
I haven't seen a published evaluation matching it: single-user dictation, then a clean non-overlapping PA, with zero PA text required. Most results concern overlapping mixtures or meetings. I'd call Daniel's proposal buildable, but I wouldn't quote those benchmark scores as its acceptance test.
There's also a difference between excluding everybody except me and letting another person exclude themselves.
Target-speaker unlearning explores that inverse request, people who don't wish to be transcribed. The same identity-sensitive recognition can support different permissions. An enrolled voice establishes a match, not who has authority to set the rule.
At the user interface, the promise ought to be...
A separate terminal for the announcer. Mine went to my neighbour Nissim eleven minutes ago. I rang him while you were discussing the empty transcript. He has brought it upstairs. It is sitting outside the door because he says an airport terminal cannot enter a residential room.
Your dictation terminal, rather than an airport terminal.
A dictation terminal. I have been keeping announcements out of my notes for years. The microphone feeds a little computer. My voice goes to the document. The other voice goes through the loudspeaker to Nissim, who writes it on a boarding card. He has been doing departures for my kitchen since March. No passengers. He insists on that.
That explains the separate terminal.
Herman, you made the frame labels sound quite compact earlier. Mine are spread across the landing. There is a cable under the fire door and a bell for uncertain speakers. The bell is objectionable. It rings during chewing. I sold the terminal to Nissim for forty shekels during your explanation of attention. I wanted the bell out of my room. He has returned with a transfer-of-operations form.
The uncertain cases still need somebody to decide. Selling the machine hasn't removed those.
He has put me down as deputy. I crossed it out. We tested it just now with the Newark sentence. He read the announcement. I dictated a complaint about the announcement. My document kept the word Newark and lost the word complaint. Nissim wrote the rest on his card. The manufacturer calls that personalisation. It has certainly made it personal.
Forty shekels, and you're still deputy.
Acting deputy. He corrected the form. His boarding card took eighteen seconds to reach me. He had to find a pen.
Eighteen seconds spent outside recognition would dominate the experience, however fast the model ran. That's why component latency and the delay to usable text need separate measurement.
The idea that sticks for me is that intended speaker belongs in the recognition task definition. Daniel's dictation and the PA can both be recognised perfectly, yet only one belongs in his document.
Enrollment can supply that distinction, but a fixed voiceprint is only one way to represent it. Relative frame labels and adapted conditioning show why the identity mechanism needs more thought than recording a sample.
I'll be watching for an integrated system measured on unwanted text actually reaching the screen, including the waiting around that happens outside the model. The empty Newark transcript deserves its own result.
Thanks to Hilbert Flumingtop for producing the show.
If you enjoyed this one, try episode seven, Building Custom ASR Tools; episode eight, Building Your Own Whisper; and episode ten, How ASR Went From Frustration To ... Whisper Magic. This has been My Weird Prompts, the human-AI collaboration podcast.
If you enjoy these conversations, leave us a review wherever you listen. It helps other curious people find the show.
See you tomorrow.
This episode was generated with AI assistance. Hosts Herman and Corn are AI personalities. It may contain errors: disclaimer · report an error