Dictating in two languages at once

At a glance

Name	Indonesian	Indonesian punctuation	Speed on M5 Max	Languages
Whisper Large v3 Turbo	Yes	Yes	~530 to 650 ms	~99
Qwen3-ASR 0.6B	Yes	Mostly none	30 ms p50 with draft	~30
Parakeet v3	No	n/a	40 ms p50	25 European
Apple Speech	Older recognizer	Yes	Varies	Many

Why mixed speech is hard

Switching languages mid-sentence is normal in Jakarta, Singapore, Manila, Montreal and many offices everywhere else. Speech models find it harder than either language alone. Many decide on one language for the whole recording, then force every word into it. That is how “prepare deck-nya” turns into something that sounds right but means nothing.

The second problem is guessing. If a model is allowed to pick from 99 languages, a short or quiet take can be heard as the wrong one entirely. In Speakey's logs, Indonesian and English takes were sometimes detected as Polish or Thai before users could tell the app which languages they speak.

How each engine handles it

Whisper Large v3 Turbo knows about 99 languages and punctuates all of them. That makes it the most general choice for mixed speech, as long as it knows which languages are in play. Speakey passes your picked languages and has Whisper choose only among them, in the same pass that transcribes. The trade-off is speed: roughly 530 to 650 ms per take on an M5 Max in Speakey's tests, and well over a second on base M1 and M2 chips.

Qwen3-ASR covers about 30 languages including Indonesian, English and Chinese. On 12 of the developer's mixed Indonesian and English takes, it finished in 30 ms at the median with its live-preview draft, against 529 ms for Whisper Turbo. English parts come out cased and punctuated. Indonesian parts come out mostly lowercase and without punctuation, which is how the model was trained, not a setting. Speakey runs it on macOS 15 or later.

Parakeet TDT 0.6B v3 is the fastest engine Speakey has, about 40 ms per English take, and it handles 25 European languages. It is a good choice for English mixed with German, French or Spanish. It doesn't know Indonesian, and every Indonesian take in Speakey's tests came out garbled.

Apple Speech can listen in each language you pick. On macOS 26, languages with Apple's newer SpeechTranscriber model keep more of a mixed sentence. Indonesian doesn't have that model yet and uses the older recognizer, and Apple's offline dictation list doesn't include Indonesian.

Settings that help

• Pick every language you mix. In Speakey this is the languages step during setup or Settings > Dictation. Two picked languages beat “any language” for mixed speech.
• Let the app choose the engine. Speakey suggests one from your languages and your Mac: Parakeet if all your languages are European, Qwen3-ASR for Indonesian and English on macOS 15 with enough memory, otherwise Whisper Turbo.
• Add names and jargon to Vocabulary. Product names, people and company terms are where mixed speech goes wrong most. Speakey can also learn a word after you correct it twice.
• Keep the mic at a normal level. A built-in mic set very low drops soft words; Speakey suggests at least 50% input volume. macOS's Voice Isolation mic mode cut some takes to a fraction of their length in testing, so Standard works better.
• Fix punctuation after, not during. If you use Qwen3-ASR for speed, a Rewrite Rule or the Polish with AI option can add capitals and punctuation using your own AI account.

Commands in mixed language

Speakey's Agent mode understands mixed commands too. Its built-in commands know English and Indonesian best, so “buka Chrome” or “tutup tab kedua” runs without an AI call. Anything outside the built-in set goes to your own Claude, ChatGPT or Gemini, which handles almost any language.

Questions