17slever17/Qwen3-ForcedAligner-Ultra

<p align="center"> <img src="https://huggingface.co/17slever17/Qwen3-ForcedAligner-Ultra/resolve/main/assets/qwen3-forced-aligner-ultra-banner.png" alt="Qwen3-ForcedAligner-Ultra" width="100%"> </p>

Qwen3-ForcedAligner-Ultra

Qwen3-ForcedAligner-Ultra is a multilingual forced aligner derived from Qwen/Qwen3-ForcedAligner-0.6B. It is designed for precise word timestamps on heterogeneous conversational, accented, noisy, and code-switched speech, even when the supplied transcript is not a literal match for the audio.

An optional small Local Head refines boundaries inside the selected acoustic island at 10 ms resolution. This addresses the inherent 80 ms timestamp-bin resolution of Qwen: the main model can select the correct acoustic island but cannot physically place every boundary with finer precision on that grid. The main Qwen model handles global word-to-audio alignment; the Local Head then refines the boundary inside a bounded +/-60 ms neighborhood.

Highlights

• —Large reduction in alignment outliers: the averaged model reduces MAE from 124.95 ms to 34.95 ms and >1000 ms errors from 1.565% to 0.195% relative to the official Qwen checkpoint on the same evaluation.
• —Precise local boundaries: the optional Local Head lowers MAE to 26.59 ms, with 78.32% of endpoints within 25 ms.
• —Transcript mismatch robustness: training includes real ASR substitutions, insertions and split/merge patterns, plus explicit missing-word examples.
• —Acoustic robustness: the mixture includes controlled speed, noise, reverberation and gain changes while retaining exact reference timestamps.
• —Multilingual and code-switched: the model retains support for all 11 Qwen3 Forced Aligner languages and was fine-tuned with explicit mixed-language utterances.
• —Drop-in base mode: without the Local Head, inference uses the normal qwen-asr forced-aligner API.

Intended use

The model is intended for word-level forced alignment when both audio and a transcript are available. Typical applications include subtitle timing, transcript repair, speech-dataset preparation, dubbing tools, and timestamping ASR output that contains occasional substitutions, insertions, or omissions.

Supported languages are Chinese, English, Cantonese, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. Chinese, English, French, German, Italian, Japanese, Korean, Russian, and Spanish received the most direct fine-tuning and evaluation coverage; Cantonese and Portuguese retain the underlying Qwen support and should be validated for the target domain.

Benchmark results

All accuracy results use one recording-disjoint multilingual evaluation:

• —6.5h of audio;
• —105,022 start/end timestamp endpoints;
• —nine languages;
• —exact manually annotated or manually corrected word boundaries;
• —identical audio, transcript, endpoint identities, and metric definitions.

FIXED means the official Qwen fix_timestamp() post-processing was applied once. MMS and WhisperX use their native word boundaries and do not receive Qwen-specific post-processing. Lower is better for error metrics; higher is better for accuracy and RTFx.

Alignment accuracy [%]

Boundary error and catastrophic tail

External noisy conversational test

One additional external test used more difficult, less clean conversational audio outside the main multilingual evaluation. It shows that the Qwen-based models transfer substantially better to this recording style, especially in the catastrophic tail. This is a small start-boundary test, so it is presented as supporting evidence rather than the primary benchmark.

Speed on NVIDIA GeForce RTX 5070 Ti

Training overview

Qwen3-ForcedAligner-Ultra starts from the official Qwen/Qwen3-ForcedAligner-0.6B checkpoint and uses a conservative multilingual LoRA fine-tune with a trainable timestamp head. Three neighboring merged checkpoints are averaged to reduce run-specific variance.

The selected training mixture contains approximately 70 hours of diverse, multilingual speech. Its clean supervision uses high-quality manual or manually corrected word boundaries. Derived examples preserve those gold timestamps while changing the input conditions:

• —Transcript deletions: spoken words remain in the audio but are omitted from the transcript, teaching surrounding words not to drift.
• —Real ASR errors: ASR substitutions, insertions, repetitions, and split/merge errors are aligned back to the original gold speech. Unmatched hallucinated words are masked instead of receiving fabricated timestamps.
• —Acoustic transforms: speed, noise, reverberation, and gain variations retain the exact word-boundary supervision.
• —Code switching: mixed-language speech and transcripts are included explicitly rather than inferred from monolingual examples.

The data mixture was balanced for linguistic and acoustic diversity rather than optimized for a single benchmark or recording style.

Usage without Local Head

Install the official Qwen runtime:

pip install -U qwen-asr

import torch
from qwen_asr import Qwen3ForcedAligner

aligner = Qwen3ForcedAligner.from_pretrained(
    "17slever17/Qwen3-ForcedAligner-Ultra",
    dtype=torch.bfloat16,
    device_map="cuda:0",
)

results = aligner.align(
    audio="speech.wav",
    text="The transcript to align.",
    language="English",
)

for word in results[0]:
    print(word.text, word.start_time, word.end_time)

Audio may also be passed as a URL, base64 payload, local path, or (numpy_array, sample_rate) tuple. Batch inference follows the normal qwen-asr API.

Usage with Local Head

Install the official Qwen runtime and the small companion package:

pip install -U "https://huggingface.co/17slever17/Qwen3-ForcedAligner-Ultra/resolve/main/companion/dist/qwen_fa_ultra-0.1.1-py3-none-any.whl"

import torch
from qwen_fa_ultra import LocalBoundaryAligner

aligner = LocalBoundaryAligner.from_pretrained(
    "17slever17/Qwen3-ForcedAligner-Ultra",
    device="cuda:0",
    dtype=torch.bfloat16,
    head_batch_size=512,
)

results = aligner.align(
    audio="audio.wav",
    text="I really thought that was funny",
    language="English",
)

for word in results[0]:
    print(word.text, word.start_time, word.end_time)

Batch inference accepts lists for all three inputs:

results = aligner.align(
    audio=["english.wav", "french.wav"],
    text=["Good morning", "Bonjour tout le monde"],
    language=["English", "French"],
)

from_pretrained() loads the main checkpoint through the official qwen-asr runtime and resolves local_head/config.json, model.safetensors, and the persisted feature normalizer from the same Hub repository. It does not require a Transformers fork or trust_remote_code.

Internally, production inference runs Qwen exactly once, retains the timestamp hidden states from that same forward pass, and applies official fix_timestamp() once. It then extracts batched acoustic features only around the FIXED boundaries and predicts bounded +/-60 ms corrections. If a pair of corrections would make a word shorter than 10 ms, the two boundaries are minimally projected around their shared midpoint to preserve a valid interval. Neighboring words are then projected onto a strictly non-overlapping timeline. When two words overlap, the runtime trusts the more accurate next-word START and trims the previous word END to that boundary. Only when this would make the previous word shorter than 10 ms is the shared boundary moved forward by the minimum amount needed to preserve valid durations. RAW Local Head features are not computed in production.

The Local Head remains optional. Loading the repository with Qwen3ForcedAligner.from_pretrained() instead uses Ultra through the unchanged official API and skips local refinement entirely.

Citation

@misc{17slever17_qwen3_forced_aligner_ultra_2026,
  author       = {17slever17},
  title        = {Qwen3-ForcedAligner-Ultra},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/17slever17/Qwen3-ForcedAligner-Ultra}}
}