Qwen3-ASR Pipeline — Parameter Guide

Controls how the speech segmenter (VAD) detects speech vs. silence. The preset automatically adjusts detection thresholds, padding, and duration filters for the selected segmenter backend. This is the single most impactful setting for capturing or ignoring quiet speech.

Preset	Behaviour	Best for
Aggressive	Low threshold (0.2), tight padding (150ms), keeps very short utterances (50ms+). Detects speech early — less padding needed since boundaries are already accurate. Captures soft whispers, faint moaning, trailing particles.	Noisy audio, soft speech, whispered dialogue, maximum subtitle coverage
Balanced	Default thresholds (0.35), moderate padding (250ms), 100ms min speech. Good tradeoff between capturing speech and avoiding false positives.	Most JAV content, general use
Conservative	High threshold (0.5), compensating padding (350ms), 150ms min speech. Detects speech later — extra padding recovers the clipped edges. Only captures clear, confident speech.	Clean studio audio, or when you want fewer subtitles (skip background chatter)

Sensitivity acts as a base layer. If you also set custom segmenter parameters in the Customize Parameters modal (e.g., a specific threshold value), your custom values override the matching preset values. Non-overridden preset values are kept. This lets you start from a preset and fine-tune individual parameters.

Splits the full audio into scenes (12–48s chunks) before any other processing. Each scene is processed independently through the pipeline. The method controls how boundaries are found:

Semantic — Uses audio embeddings and clustering to find natural scene transitions. Has true merge logic that guarantees the minimum scene duration is respected. Best for Qwen (default).
Auditok — Energy-based silence detection with a two-pass coarse+fine strategy. Fast and reliable; default for Whisper pipelines.
Silero — Uses the Silero VAD model for scene-level silence detection. Different from the segmenter — this runs at scene scale, not frame scale.
None — No scene splitting. Entire audio is processed as one scene. Only use for very short files (<3 minutes).

Optional audio preprocessing that runs per-scene after scene detection. Cleans up audio before it reaches the VAD segmenter and ASR model. Enhancement uses additional VRAM.

None — No enhancement. Raw audio goes directly to segmentation. Recommended for most content as Qwen3-ASR handles noise well on its own.
FFmpeg DSP — Lightweight audio filters (loudnorm, denoise) running in FFmpeg. Zero GPU cost. Good for normalizing volume levels.
ZipEnhancer (Torch) — Neural speech enhancement at 16kHz. Effective but adds ~1GB VRAM. Good for noisy recordings.
ClearVoice — Multiple denoising models (FRCRN 16kHz, MossFormerGAN 16kHz). Use FRCRN for general denoising.
BS-RoFormer — Vocal isolation (separates speech from music/background). Not yet available.

Within each scene, the segmenter runs Voice Activity Detection (VAD) to find exactly where speech occurs. Detected speech regions are grouped into frames (max 6s each) and sent to the ASR model. The sensitivity preset above controls this backend's parameters.

Silero v6.2 — Recommended. Force-splits long speech at internal silences (max_speech_duration_s), hysteresis via neg_threshold for stable segmentation in noisy audio. Handles fast Japanese dialogue well.
TEN VAD — Previous default. Lightweight energy-based VAD. Fast but can miss pauses in rapid dialogue with background audio.
Silero v4.0 / v3.1 — Older Silero versions without force-split or hysteresis. Use only if v6.2 causes issues.
Whisper VAD — Uses a Whisper model internally for VAD. Accurate but slow and VRAM-heavy. Variants: tiny, small, medium.
NeMo Lite — NVIDIA NeMo-based VAD. Requires nemo_toolkit (not included in default install).
None — Skip segmentation. Entire scene goes to ASR as one chunk. Only use with Full Scene framing.

Quick model selection in the ensemble row. The 1.7B model requires ~8GB VRAM; the 0.6B requires ~4GB. This is the same as the ASR Model setting in the Customize Parameters modal (Tab 1) — changing it in either place updates both.

The speech recognition model that converts audio to text. The 1.7B parameter model is more accurate and handles complex speech better. The 0.6B model is faster and uses roughly half the VRAM (4GB vs 8GB).

Forces the model to transcribe in a specific language instead of auto-detecting. For JAV content, forcing Japanese avoids the model occasionally switching to Chinese for similar-sounding phonemes.

Free-text hints that help the model recognize specific names and terms. The model uses this as context for improved accuracy on proper nouns. Enter actress names, studio names, or domain-specific terminology.

Where the model runs. Auto detects your GPU; CUDA forces GPU; CPU forces processor-only (much slower). Most users should leave this on Auto.

Model precision. Float16 is fastest on most NVIDIA GPUs. BFloat16 is best on Ampere+ GPUs (RTX 30xx/40xx). Float32 uses more memory and is slower but may be needed on older hardware. Auto selects the best option for your GPU.

The attention algorithm used internally. SDPA (Scaled Dot-Product Attention) is fastest on most GPUs. Flash Attention 2 requires a separate install and specific hardware. Eager is the slowest but most compatible fallback.

Controls how audio is divided into frames for text generation. Each frame is sent to the ASR model as an independent unit. This setting determines what the model “sees” at a time.

VAD Grouped — Groups VAD speech segments into frames up to Max Group Duration (default 6s). Each frame contains only detected speech regions with natural pause boundaries. Best accuracy for most content.
Full Scene — Sends the entire scene (12–48s) as a single frame. The model gets maximum audio context but may struggle with precise timing on long scenes. Can produce better results for content with continuous dialogue and few pauses.
SRT Source — Uses an existing SRT file to define frame boundaries. Useful for re-transcription or translation workflows where timing from a reference subtitle is desired.

Enforces scene boundaries so no audio segment exceeds the ForcedAligner's 180-second (3-minute) processing limit. When enabled, any scene longer than 180s is automatically re-split at silence boundaries. Disabling this risks aligner failures on long scenes.

Minimum length for a detected scene. Scenes shorter than this are merged with their neighbors. Too low and you get fragmented tiny scenes with lost context. Too high and the detector can't split at natural boundaries.

Maximum length for a detected scene. Scenes longer than this are forcibly split. The default of 48s keeps scenes well within the aligner's 180s limit while preserving context.

When using VAD Grouped framing, this sets the maximum silence gap (in seconds) allowed within a single frame. If a silence gap between two speech segments exceeds this threshold, a new frame starts. Lower values produce more, smaller frames with tighter sentence boundaries. Higher values produce fewer, larger frames with more context per frame.

When using VAD Grouped framing, this sets the maximum duration of each group. VAD speech segments within a scene are grouped together until they hit this limit, then a new group starts. Shorter groups mean more precise text generation and alignment but add processing overhead.

Speech detection probability threshold for the VAD model. A frame of audio is classified as speech when the model's confidence exceeds this value. Lower values detect quieter or more ambiguous speech (more sensitive), higher values require clearer speech signals (fewer false positives). This slider overrides the value set by the sensitivity preset.

0.05–0.15 — Very sensitive. Captures whispers and background murmuring. May produce false positives from music or ambient noise.
0.20–0.35 — Balanced range. Good for most JAV content with varying vocal intensity.
0.40–0.60 — Conservative. Only clear speech is detected. Quiet moans or whispered dialogue may be missed.
0.65–0.80 — Very strict. Only loud, clear speech passes. Use for extremely noisy source material.

Milliseconds of audio added before and after each detected speech segment. Padding ensures the ASR model sees the full onset and release of speech, preventing clipped words. This slider overrides the value set by the sensitivity preset.

50–150ms — Tight padding. Use with aggressive sensitivity (which already detects speech early) to avoid excessive silence in frames.
200–300ms — Standard range. Good default for balanced sensitivity.
350–600ms — Wide padding. Use with conservative sensitivity (which detects speech late) to capture soft onsets and trailing particles.

The sensitivity dropdown sets a complete preset of VAD parameters. The VAD Threshold and VAD Padding sliders override only the specific values you change, leaving other preset parameters (like neg_threshold, min_speech_duration_ms) intact. If you don't touch the sliders, the sensitivity preset applies fully.

How many audio frames are processed simultaneously by the ASR model. Batch size 1 processes frames one at a time, giving the model full attention and best accuracy. Higher values use more VRAM but process faster.

Maximum number of text tokens the model can generate per frame. 4096 tokens covers roughly 5-10 minutes of spoken audio. This is a safety ceiling, not a target — most frames use far fewer tokens.

Penalizes the model for repeating the same tokens. A value of 1.0 disables the penalty entirely. Values above 1.0 make repetition progressively less likely. Repetition manifests as the model generating the same word or phrase in a loop (e.g., あああああ or the same sentence repeated).

Maximum tokens the model is allowed to generate per second of audio. This is a safety net that stops runaway generation — if the model hallucinates, it starts producing far more text than real speech warrants. Normal Japanese speech produces roughly 5-10 tokens per second.

The ForcedAligner takes the generated text and the original audio, then produces precise word-level timestamps. This is what makes subtitles appear at the right moment. Disabling it (None) falls back to VAD-based timing, which is less precise. When set to None, the aligner model is not loaded at all, saving ~2GB VRAM.

The model used for forced alignment. Currently only the 0.6B variant is available. It has a hard processing limit of 180 seconds per segment (enforced by Safe Chunking in the Audio tab).

Cleans up ASR output before alignment. Removes artifacts like stray punctuation, repeated filler sequences, and formatting issues that would confuse the aligner. Passthrough skips all cleaning and sends raw ASR output directly to alignment.

How word timestamps are resolved from the aligner's output. The four modes offer different tradeoffs between precision and coverage:

Aligner + VAD Fallback (recommended) — Uses aligner timestamps when available; for words the aligner couldn't place, falls back to VAD segment boundaries with speech-aware gap filling. Best overall accuracy and coverage.
Aligner + Interpolation — Uses aligner timestamps, fills gaps by interpolating evenly between placed words. Better for sparse aligner output where VAD regions aren't available.
Aligner Only — Uses raw aligner timestamps with no recovery. If the aligner collapses, subtitles may have timing issues. Useful for studying raw aligner quality.
VAD Only — Skips the aligner entirely (not loaded, saves VRAM). All word timestamps come from VAD frame boundaries. Fastest but least precise. Use when aligner is disabled or VRAM is critical.

When alignment fails for a scene (all word timestamps collapse to a single point), step-down automatically retries the scene with tighter framing — breaking the audio into smaller segments and re-running generation + alignment. This recovers most collapsed scenes without user intervention. Hidden when timestamp mode is VAD Only (not applicable).

The initial group duration used when step-down is triggered. The scene is re-framed into segments of this length and re-processed. If this tier also fails, Tier 2 is attempted.

The fallback group duration if Tier 1 step-down also fails. This is the tightest framing before giving up on alignment for a scene. Smaller values give the aligner less audio context but are more likely to avoid collapse.

Controls how the raw word-level output from the aligner is grouped into subtitle lines. This is independent of the Post-processing Preset below, which fine-tunes gap thresholds and merge behavior within the selected regrouping mode.

Standard — Full regrouping: gap-based splitting, fragment merging, punctuation-based splitting, and duration/character safety caps. Best readability for most content. Branch A uses REGROUP_JAV; Branch B uses text-only splitting (no gap heuristics, since VAD-only timestamps are synthetic).
Sentence Only — Text-only splitting: punctuation-based splitting and safety caps (80 characters, 8 seconds) without gap heuristics. Use when gap-based splitting causes problems like merging separate lines or splitting mid-thought. Natural sentence boundaries are preserved.
Off (Recommended) — Frame-native output: each temporal frame becomes one subtitle line, preserving natural speech segmentation boundaries. No regrouping, merging, or splitting is applied. Best overall quality in testing.

Controls how subtitle lines are regrouped and formatted in the final SRT output. Each preset tunes gap-splitting thresholds, maximum subtitle duration, and merge behavior:

High Moan — Optimized for JAV content. Splits aggressively at short gaps (1.5s), caps subtitle lines at 8 seconds, and preserves breathy/moaning segments as separate subtitles.
Default — General-purpose regrouping. Standard gap thresholds and merge behavior suitable for most spoken content.
Narrative — Optimized for story-driven content with longer dialogue. Allows longer subtitle lines and merges more aggressively to maintain sentence flow.