Taraneh-Q0.6B: Persian Automatic Speech Recognition

Taraneh-Q0.6B is the Qwen3-ASR-0.6B member of Taraneh, a family of Persian ASR models fine-tuned on about 1,560 hours of speech. The base checkpoint is Qwen/Qwen3-ASR-0.6B.

License

Apache-2.0, the same license as Qwen3-ASR. Commercial use is allowed, including in products and hosted services. Apache-2.0 includes a patent grant. Keep this notice and the original Qwen license notice with the weights.

Paper

Aref Farhadipour. Taraneh-ASR: Scaling Persian Speech Recognition and Forced Alignment Across Singing and Conversational Domains. University of Zurich.

• PDF in the code repository: https://github.com/areffarhadi/persian-asr/blob/main/ArefFarhadi-TaranehASR.pdf
• Code, manifests, normalization, and inference: https://github.com/areffarhadi/persian-asr

Load this checkpoint with the qwen-asr package. Offline and streaming inference use vLLM 0.14 (pip install "qwen-asr[vllm]"). Streaming is available only with the vLLM backend.

Forced alignment

This checkpoint can return word times as well as a transcript. In the GitHub scripts, --task asr writes the transcript and --task asr+fa also writes alignments.jsonl.

The readout uses thinker self-attention heads L22/H7, L24/H12, and L20/H15, a median filter of width 3, column normalization, dynamic time warping, and a −35 dB trim of the first-word start and last-word end. On the FarsDat gold set this model reaches 47.2 ms mean absolute error. Heads and the command are in the GitHub README.

Results

Word error rate / character error rate (%) after the unified Persian text normalization. PSRB and VisualEars were unseen in training.

FLEURS	Common Voice	YouTube	Music	Media	PSRB	VisualEars
7.11 / 2.54	9.65 / 3.86	18.87 / 10.12	35.97 / 15.83	22.32 / 11.73	24.61 / 10.67	9.44 / 3.12

The full comparison of all five Taraneh models is in the paper and in the GitHub README.