Does Wan 2.2 Have Audio? Silent Clips, S2V Lip Sync and Free Fixes (2026)

The first time I delivered a Wan 2.2 clip to a client, I sent the file, waited, and got back: "Is the sound missing or is my player broken?" Neither. Wan 2.2's main checkpoints are silent models — they were never trained to output an audio track — and nothing in the interface warns you about it, because "video without sound" is not an error state, it is just a video.

I have since set up the workflow that fixes this properly, and it splits into three paths depending on what you actually need: a talking head, a scored brand clip, or a sound-effect pass on an otherwise finished shot. This guide covers which Wan 2.2 checkpoints can produce audio at all, why the model is silent by default, and the three ways to end up with a file that has sound — including one that costs nothing.

TL;DR

• Wan 2.2's base checkpoints do not generate audio. T2V-A14B, I2V-A14B and TI2V-5B all output silent video. This is architectural, not a settings mistake you can toggle.
• Audio enters through S2V-14B, a separate checkpoint that takes an audio track plus an image and drives a talking or singing performance from it. It consumes audio; it does not create it.
• Native audio generation started with Wan 2.5. If you want the model to invent sound effects, ambience or dialogue, you need 2.5 or newer — 2.6 and 2.7 both build on it.
• The free fix is a post pass. Generate silent, then add voice with a TTS tool and a music bed, and sync on the beat. For most social and product work this is faster than switching models.
• Check the file before you blame the player. Wan outputs H.264 video with no audio stream at all in the silent checkpoints, so some players show no waveform rather than a muted one.

Generate a Wan 2.2 Clip Free →

The short answer, and the one that surprises people

Wan version	Native audio?	How audio enters
Wan 2.1	No	Post-production only
Wan 2.2 (T2V / I2V / TI2V)	No	Post-production only
Wan 2.2 S2V-14B	No — it needs audio input	You supply a track; it drives lip sync
Wan 2.2 Animate-14B	No	Post-production only
Wan 2.5	Yes	Generated alongside video
Wan 2.6	Yes	Generated, plus voice reference
Wan 2.7	Yes	Generated, plus voice reference

The confusing row is S2V-14B, because it is usually described as the "audio" variant of Wan 2.2 — and it is, in the sense that audio is its input. It takes a photo plus a speech or singing track and produces a video where the mouth and head move in time with that track. It does not synthesise the track. Load it without audio and you get an error, not a silent clip.

Why Wan 2.2 is silent

Wan 2.2 is a diffusion video model trained on visual data. Its five checkpoints are specialised on different visual tasks — text-to-video, image-to-video, compact hybrid, speech-driven, motion transfer — but audio generation is not among those tasks for any of them. The model has no audio decoder, so there is no latent representation of sound to produce.

This is a general property of open video models of that generation, not a Wan-specific limitation. What changed in Wan 2.5 was the addition of an audio branch that generates synchronised sound effects, ambience and speech alongside the visual stream. That is a genuinely different architecture, which is why 2.5 and later cannot be replicated by prompting 2.2 harder.

Practically: if a platform advertises "Wan 2.2 with sound," one of two things is happening. Either it is running an S2V job and you supplied the audio, or it is stitching Wan 2.2 video to a separately generated audio track and calling the result one model. The second is fine, but know what you are buying.

If you want output today, start here: Launch Wan 2.2 Now →

Route 1: S2V-14B for talking heads

If your shot is a person speaking to camera, S2V-14B is the right tool and it produces the best lip sync in the open Wan family.

The workflow:

1. Prepare the audio first. A clean mono or stereo track, 16-bit, no music bed underneath the speech. S2V reads the waveform; background music confuses the alignment.
2. Choose a front-facing image. Face to camera, mouth visible, no hands over the mouth, even lighting. Profile shots produce weak sync because the model has less mouth geometry to work with.
3. Keep the frame tight. A medium close-up gives the model enough facial detail. Full-body shots at 720p leave about 40 pixels of mouth, which is not enough to animate convincingly.
4. Match clip length to audio length. S2V generates for the duration you ask; if your audio runs 12 seconds and your clip is 5, the last seven seconds of speech have nowhere to go.

Two details that improved my results noticeably. First, trim silence at the head of the track — leading silence makes the model guess when the mouth should start, and it usually guesses early. Second, keep the head still in the prompt. If you ask for the character to turn or walk while speaking, the sync degrades because the model is now solving two motion problems at once.

For the full setup including image preparation and what CosyVoice integration adds for speech synthesis, the talking avatar guide walks through it end to end. And if you are choosing between Wan's speech-to-video approach and dedicated avatar platforms, this comparison covers where each one breaks.

Route 2: use a Wan version that generates audio

If you want the model to invent the sound — waves crashing, a door slamming, ambient street noise — you need Wan 2.5, 2.6 or 2.7, hosted. None of those are available as open downloads, so this route means a browser tool or API rather than a local install.

What native audio actually gives you:

• Synchronised sound effects. The clink of a glass lands on the frame where the glass moves, because audio and video are generated together rather than assembled afterwards.
• Ambience that matches the scene. A rooftop shot gets wind and city hum; an interior gets room tone. This is the part that post-production most often fakes badly.
• Speech when the model supports it, with 2.6 and 2.7 adding voice reference so a character's voice stays consistent across clips.

What it still does not give you: music. Native audio in the Wan family is oriented toward effects, ambience and speech, not scored music. Plan a music bed separately regardless of which model you use.

Try a Wan Model With Native Audio →

Cost is the trade. Wan 2.2 is free to run locally once you own the hardware; native-audio Wan versions are metered. For a 15-second 1080p clip on current hosted Wan 2.7 pricing, you are paying per generation whether or not the audio is usable — and roughly a third of the time it is not, because the model invents an effect you did not want. Budget for two or three takes per usable clip.

Route 3: the free post pass

This is what I do most often, and it is the cheapest reliable path: generate silent Wan 2.2 video, then build the audio in post from three layers.

Voice. For narration, a text-to-speech tool is enough. Qwen3 TTS is free and handles Chinese and English well; for short English voiceovers I also use the built-in speech in an editing app. Generate the voice first, then cut the video to the voice — never the other way around, because stretching video to fit audio produces speed ramps that look wrong.

Effects. Two or three sound effects do more for perceived quality than twenty. The ones worth buying or sourcing: a room-tone bed (constant low-level ambience that makes silence stop sounding like a failed export), a whoosh for transitions, and one or two subject-appropriate effects. Free libraries cover all three.

Music. One track, ducked under the voice to about -18dB. The most common mistake in AI video is music at full level competing with narration, which makes both unintelligible.

Ready to try it yourself? Try Wan 2.2 Free →

The whole pass takes about ten minutes for a 15-second clip and requires no model changes. For a longer walk-through of assembling a scored clip from silent generations, the free AI video generator with sound guide covers the layering order.

Which route for which job

Your shot	Best route	Why
Person talking to camera	S2V-14B	Best lip sync in the open family, and you already have the voice
Product clip with effects	Wan 2.7 hosted	Effects land on the right frame
Social clip with music	Silent 2.2 + post	Fastest, free, full control of the bed
Brand film with ambience	Wan 2.6 / 2.7	Ambience generation is the real strength
Loop or abstract motion	Silent 2.2 + music	No sync problem to solve
Dialogue between two characters	Wan 2.7 + post	Native speech plus a second voice track in post
For a closer look at how it stacks up against other models, see Wan Video Models Compared.
If you want to test it without installing anything, the Wan 2.2 Animate tool works in the browser.

Where the audio workflow goes wrong

Symptom	Cause	Fix
No waveform at all in the player	Silent checkpoint — no audio stream	Expected; audio is a separate pass
Lip sync drifts late	Audio longer than the clip, or silence at the head	Trim leading silence, match durations
Mouth moves to the music	Music bed fed into S2V instead of the voice	Isolate the vocal track before generating
Voice sounds robotic	TTS default settings	Slow the rate slightly, add a short pause between sentences
Effects feel pasted on	Effect is at the wrong level or starts a frame late	Nudge start time, lower level 3–6dB under the picture
Audio crackles at the cut	Sample-rate mismatch between layers	Convert everything to 48kHz before assembling

The last one costs people the most time. I once spent forty minutes hunting a "bad model output" crackle that turned out to be a 44.1kHz music file mixed with a 48kHz voice track in a 48kHz project. Convert first, mix second.

Does Wan 2.2 have audio if I use it in the browser?

Depends which model the browser tool is serving. A free tool running the open Wan 2.2 checkpoints produces silent clips, exactly as a local install does — the model is the same, only the hosting differs. A tool offering "Wan 2.7" or "Wan 2.5" produces native audio because those versions generate it.

The practical check before you generate: look for the model version in the interface, not the site name. If it says 2.2, plan the post pass. If it says 2.5, 2.6 or 2.7 and has an audio toggle, you can skip it.

Generate With Audio on Wan 2.7 →

The Bottom Line

Wan 2.2 does not have audio. Its text-to-video, image-to-video and hybrid checkpoints output silent video, and the one variant with audio in its name — S2V-14B — takes an audio track as input rather than producing one. This is not a bug to work around with better prompts; it is what the model is.

Three routes fix it. S2V-14B if you have the voice track and need lip sync. Wan 2.5, 2.6 or 2.7 if you want the model to invent effects and ambience. A free post pass with a TTS voice, two effects and a ducked music bed if you want speed and control — which, for social and product work, is usually the answer.

If you just want to see a Wan 2.2 clip and decide where the sound should go, generate one free in the browser and build the audio against it. For longer shots where native audio saves the whole pass, Wan 2.7 on Pollo is the hosted route I keep as a fallback.

FAQ

Does Wan 2.2 have audio?

No. Wan 2.2's main checkpoints — T2V-A14B, I2V-A14B and TI2V-5B — generate silent video with no audio stream. The S2V-14B variant accepts an audio track and drives lip sync from it, but it does not generate audio itself. Native audio generation begins with Wan 2.5.

Why is my Wan 2.2 video silent?

Because silent output is the default and expected behaviour for those checkpoints. Some players show no waveform rather than a muted one, since the file has no audio stream at all. Adding sound requires either the S2V checkpoint with a supplied track, a newer Wan version, or a post-production pass.

Can Wan 2.2 generate lip sync?

Yes, through S2V-14B. It takes an image plus a speech or singing track and produces video where the mouth and head move in time with that audio. Sync quality depends heavily on the source image — front-facing, evenly lit, mouth unobstructed, framed no wider than a medium close-up.

Which Wan model generates audio automatically?

Wan 2.5 was the first to generate audio alongside video, and Wan 2.6 and 2.7 build on it with voice references and better ambience. All three are hosted-only; none has published open weights, so access is through browser tools or the API.

Can I add audio to Wan 2.2 video for free?

Yes, and this is the most common workflow: generate the clip silently, then add a TTS narration track, two or three sound effects and a music bed in any editing app. Total time is about ten minutes for a 15-second clip, and the result is usually more controllable than letting a model invent audio.

Does Wan 2.2 S2V need a specific audio format?

Use a clean 48kHz track with the speech isolated — no music underneath. Leading silence should be trimmed so the model starts the mouth at the right frame. Matching the clip duration to the audio duration avoids sync drift at the end of the shot.

References

• Wan2.2 official repository — checkpoint list including S2V-14B, architecture notes and Apache 2.0 licence terms.
• Wan2.2-S2V-14B on Hugging Face — speech-driven video model card, audio input requirements and CosyVoice integration notes.
• Wan2.2-TI2V-5B on Hugging Face — compact model reference confirming 720p 24fps silent output.
• Wan official site — hosted Wan family playground, including the versions that generate audio.
• ComfyUI Wan 2.2 tutorial — node setup, including the separate audio path used with S2V.

Related guides

• Wan 2.2 AI video generator: the four variants explained — which checkpoint to load for which job.
• Wan 2.2 S2V talking avatar guide — the speech-driven workflow in detail.
• Wan AI speech-to-video vs talking avatar generators — where Wan's approach wins and where it doesn't.
• Free AI video generator with sound — layering voice, effects and music on a silent generation.
• Qwen3 TTS free text to speech — the free voice layer for the post pass.

Start With a Free Wan 2.2 Clip →

Want Native Audio? Try Wan 2.7 on Pollo →