The ASR is solved. The grader is not.

As of May 2026, no. The closest add-ons in the AwesomeTTS / HyperTTS family only add text-to-speech to the card front. The open-source anki-voice (github.com/williamknows/anki-voice) is voice navigation: it routes spoken commands like 'again', 'good', and 'easy' to the Anki review buttons via AnkiConnect, but never compares your spoken answer to the back of the card. A long-running feature request on Anki-Android (issue #1754) asks for spoken-answer grading and is still open. If you want spoken-answer grading today, the working options are outside Anki: Audio Flash (audioflash.app, imports .apkg, FSRS) and the AI Voice Tutor for Anki shared on forums.ankiweb.net in late 2025 by user albertjoseph.

The grader is the component that decides whether your answer was correct. For a multiple-choice card the grader is trivial: did you pick the right letter. For a free-response card the grader has to read what you said, read what the card lists as the correct answer, and decide if those match in substance. Substring overlap is not the right metric. 'The renal artery branches off the abdominal aorta below the SMA' and 'it branches from the abdominal aorta, below the SMA, into segmental arteries' should both be marked correct. 'The renal artery branches off the inferior vena cava' should be marked wrong. The hard part is that the model needs the actual reference material to make that call, not just the card back, because card backs are usually shorthand.

Because as of 2026 you have several production-grade options that do better than human transcription on most domain audio. Whisper-large-v3 hits sub-5 percent word-error-rate on medical lecture audio. Vosk runs offline on the laptop and handles command-style utterances well enough that the open-source anki-voice add-on never needs cloud APIs. Browser-native SpeechRecognition is fine for short answers in quiet rooms. The transcription step is no longer where the workflow breaks. The grading step is.

Audio Flash describes its grading as 'advanced voice recognition technology' that 'compares them to the card's correct answer' and is 'designed to understand natural speech patterns.' That phrasing is consistent with an LLM-as-judge approach (read the user transcript, read the card back, output a verdict), but the company does not publish a held-out evaluation, a confusion matrix, or a sample of cards where the grader gets it right and wrong. The app does import .apkg and use FSRS, which is the right scheduling primitive. The grader quality is just opaque, which is why it does not show up in a leaderboard.

That tool (shared in late 2025 by user albertjoseph) ingests an .apkg, runs a conversational voice session where the AI asks questions and probes follow-ups, and downloads an updated deck preserving FSRS state. It is more ambitious than Audio Flash in that it allows back-and-forth instead of a single utterance per card. The grader is still undocumented. Without a held-out eval there is no way to know whether 'probes deeper on close answers' means the model has a rubric or is just trained to keep talking.

Because the grader is the input to the scheduler. FSRS and SM-2 both decide when to show you a card next based on whether you said you got it right. If the grader marks correct answers as wrong, you re-see those cards days earlier than needed (false positives on the 'forgot' side, study time wasted). If the grader marks wrong answers as right, the card disappears from your queue and you never relearn the fact (the worse failure mode, you carry a wrong fact into the exam). Voice rehearsal with a 60-percent grader is worse than typed rehearsal with a 90-percent grader, even when the ASR is perfect, because the bad grader poisons the schedule.

Not as a first-class feature. The product is built around four written question formats (MCQ, free-response, case-style, image-occlusion) generated from your professor's actual slide deck, PDF, textbook, or YouTube lecture. The grader is what is differentiated: 81.3 on a held-out three-document eval scored on factual correctness, clarity, distractor quality, and question-type coverage, versus a field average of 67.9 (methodology and per-criterion scores at studyly.io/quality). If you want voice rehearsal today the honest workflow is OS dictation into the free-response field, which the recipe section below walks through.

Yes. Studyly exports .apkg files including image-occlusion cards. From there you can review them in Anki proper, or import the .apkg into Audio Flash or the AI Voice Tutor for the voice loop. The reason to generate in Studyly first is the in-flight rubric: cards are gated on source-anchoring, length-matching, filler-template ban, and grammar parallelism before they are emitted (full breakdown at studyly.io/t/anki-rubric-mcq-quality). A voice grader downstream is only as good as the cards it grades; bad cards stay bad regardless of the I/O modality.

Anything modern. Built-in macOS dictation (System Settings -> Keyboard -> Dictation) handles short medical-school answers at about 95 percent word accuracy in a quiet room. iOS dictation is comparable. On Windows, Win+H opens the system dictation overlay; accuracy is similar. None of these need a third-party microphone for short answers. If you are reviewing a long stack of cards in one session, a USB headset cuts background noise enough to push the ASR error rate below 2 percent, which materially reduces friction because you stop having to fix transcription typos before the grader runs.

Three reasons people reach for voice rehearsal. One, oral exam prep: med, dental, and PA programs run viva-style assessments where you have to verbalize a differential diagnosis, and typing trains the wrong muscle. Two, hand fatigue or RSI: a student who already does six hours of clinical writing per day cannot also do two hours of card typing at night. Three, dual-task tolerance: cooking, walking, driving (audio only, eyes on the road). For oral exam prep the verbalization itself is the point and any grader at all is better than none. For RSI it is an accessibility need. For dual-tasking the grader can be lighter because the user is willing to forgive false negatives.

Yes, on free-response cards specifically. A typical Anki back field is a few words to a single sentence. If a med student says 'the loop of Henle reabsorbs sodium chloride via the NKCC2 cotransporter in the thick ascending limb' and the card back reads 'NKCC2', a string-match grader fails. A back-only LLM-as-judge does better but still confuses correct elaborations with hallucinations because it has no other source to verify the elaboration against. A grader that has the original lecture slide can mark the elaboration correct if the slide supports it, wrong if the slide contradicts it. The reference material does most of the work.

studyly.io/quality. The page lists the three held-out documents (a cardiology lecture, a renal lecture, a USMLE-style passage), the four criteria (factual correctness, clarity, distractor quality, question-type coverage), and the per-tool scores: Studyly 81.3, Unattle 78.0, Gauntlet 68.0, Turbolearn 57.8. Audio Flash and the AI Voice Tutor for Anki are not on the leaderboard because we have no public eval to compare against; both are welcome to publish one.