voice_audio_multi/docs/models/qwen3_forced_aligner.md
Fabio c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
Initial import: audio.cpp with Orange Pi documentation
2026-11-04 19:05:22 +01:00

2.3 KiB

Qwen3 Forced Aligner

For shared conversion guidance, see Qwen3 GGUF checkpoints.

The forced aligner maps an exact transcript onto speech audio. It is not an ASR route: the transcript is required input. Standalone forced alignment does not chunk audio because exact transcript/audio chunk pairing cannot be inferred safely from a raw transcript. For long-audio timestamping, use Qwen3 ASR with --words-out and qwen3_asr.forced_aligner_path; ASR chunks the audio first, then aligns each recognized transcript to its matching audio chunk.

Field Value
Family qwen3_forced_aligner
GGUF model models/Qwen3-ForcedAligner-0.6B-GGUF/qwen3-forced-aligner-0.6b-q8_0.gguf
Task align
Modes offline
Input Speech WAV plus exact transcript
Output Word timestamp JSON through --words-out
audiocpp_cli --task align --family qwen3_forced_aligner --model models/Qwen3-ForcedAligner-0.6B-GGUF/qwen3-forced-aligner-0.6b-q8_0.gguf --backend cuda --audio assets/resources/sample_16k.wav --text "Some call me nature, others call me Mother Nature." --language English --words-out words.json

Common Options (use directly)

Option Values Default Meaning
--audio WAV path required Speech audio at any sample rate; timestamps are reported in that file's sample domain.
--text exact transcript required Transcript to align.
--language language code required Transcript language.
--audio-chunk-mode auto, none auto Standalone forced alignment runs one audio/transcript pair. fixed and vad are rejected because transcript chunk boundaries would be ambiguous.
--words-out JSON path not set Word timestamp output.

Request Options (use with --request-option)

Option Values Default Meaning
clamp_timestamps_to_audio bool false Opt-in guard for word timestamps: keep repaired spans inside the local audio input. Default preserves existing timestamp repair behavior.

Session Options (use with --session-option)

Option Values Default Meaning
qwen3_forced_aligner.weight_type native, f32, f16, bf16, q8_0 native Aligner thinker weight type.