Qwen: Qwen3 ASR 1.7B

Qwen3-ASR-1.7B is Qwen’s larger speech-recognition model in the Qwen3-ASR family. It takes audio and returns recognized text together with a language label. The family builds on the audio-understanding foundation of Qwen3-Omni and uses one ASR model for offline and streaming recognition. The 1.7B edition places emphasis on recognition quality, while the separately named 0.6B edition emphasizes efficiency. Those editions remain distinct choices even though they share the documented language coverage and recognition workflows.

Language identification and audio coverage

The manufacturer lists 30 languages and 22 Chinese dialects, as well as English accents from multiple regions. The language list includes Chinese, English, Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, Swedish, Danish, Finnish, Polish, Czech, Filipino, Persian, Greek, Hungarian, Macedonian and Romanian. The supported audio types include speech, singing voice and songs with background music. These capabilities make the transcript useful for interviews, lectures, voice recordings and lyric review, with the original audio retained for checking difficult passages.

Preparing audio and obtaining a transcript

The producer’s toolkit accepts an audio-file reference, encoded audio or a waveform paired with its sample rate, and its offline workflow can process a collection of recordings. It can detect the language automatically or use a specified supported language. Choose automatic detection for an unknown recording; provide the actual spoken language when it is known and relevant to the task. The example result exposes the detected or selected language and recognized text. The language setting guides recognition of the recording rather than defining a request to rewrite its content in another language.

For an interview, keep the recording associated with its date, subject and source identifier, then use the returned text as a searchable transcript. For a lecture, preserve technical terms and numerical expressions during review instead of silently replacing unclear words with a plausible explanation. For a song, check the recognized lyrics while listening to the vocal line, especially where music masks a syllable. The transcript can become input for a separate summarization, translation or editorial step once its wording has been checked against the audio.

Streaming recognition and its scope

The same ASR family supports recognition while audio arrives. The manufacturer’s streaming example feeds successive audio chunks into a retained recognition state, displays interim text, then finishes the stream to obtain the final language and transcript. Keep a distinction between the current incremental text and the completed result when preparing live captions or a running transcript. The documented streaming implementation does not combine streaming with batch inference or return timestamps; a timed transcript uses a different workflow. These are producer implementation details rather than a promise about an application’s transport or live-caption interface.

Word or character timing with a separate aligner

Qwen3-ForcedAligner-0.6B aligns an audio recording with supplied text and predicts word- or character-level start and end times. It is a separate model from both Qwen3-ASR editions. The ASR toolkit can use that aligner when timestamps are requested in its offline transcription flow, or alignment can be run on an existing text–audio pair. The aligner’s documented coverage is 11 languages: Chinese, English, Cantonese, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. Its stated speech span is up to five minutes; this is the aligner’s scope, not a five-minute ceiling for ASR transcription.

For subtitle preparation, first obtain and review the transcript, then align it to the source recording where that alignment workflow is appropriate. Use the returned start and end times to prepare caption segments and inspect their boundaries while listening. Alignment associates supplied words with audio; it should not be treated as proof that an incorrect transcript has been corrected. The useful artifact is a checked transcript with timing information that remains connected to the recording it describes.

Long recordings and practical limits

The family supports long-audio transcription, and the official examples instruct users to allow more generated text for longer input. Their illustrative token settings are configuration choices rather than a universal model context or audio-duration limit. Plan the output allowance around the recording and check that the ending is present. Review names, quantities, language changes, obscured words and overlapping speech against the original sound. Language coverage and robustness to challenging audio do not guarantee a flawless transcript for every accent, recording or musical arrangement. A complete result preserves the recognized wording, its language and its relationship to the audio so it can be searched, edited or used in a later content workflow.