Qwen3-ForcedAligner-0.6B-gguf

Qwen3-ForcedAligner-0.6B — GGUF (for transcribe.cpp)

GGUF conversion of Qwen/Qwen3-ForcedAligner-0.6B (checkpoint commit c7cbfc2) for the transcribe.cpp engine. Converted with scripts/convert-qwen3_forced_aligner.py and quantized with tools/transcribe-quantize from the Adi2K/transcribe.cpp fork, branch feat/qwen3-forced-aligner @ b37f24d.

The model is a non-autoregressive forced aligner: given audio plus its transcript and language, it emits word-level timestamps in a single pass (80 ms resolution, ~400 s hard ceiling per call; publisher validates ≤300 s). It does not transcribe on its own.

Provenance & caveats

• Independent conversion for integration testing (Isaree on-device scribe), not an official artifact of either upstream project.
• Q8_0 reproduces the bf16 Python reference (QwenLM/Qwen3-ASR qwen-asr 0.0.6) exactly on the engine's jfk fixture: 22/22 words, 0 ms timestamp deviation (Metal backend).
• Requires a transcribe.cpp build that includes the qwen3_forced_aligner family (fork branch above; not yet in upstream). Not llama.cpp-compatible — uses the stt.qwen3_forced_aligner.* GGUF KV namespace.
• ja/ko are dropped from the source model's language list in this conversion.

Usage

transcribe-cli -m Qwen3-ForcedAligner-0.6B-Q8_0.gguf \
  -l en --align-text transcript.txt audio.wav

Original model © Alibaba Cloud / Qwen team, Apache-2.0.