Rivya Audio Uploads and Recording Requirements
Upload audio for supported video-reference tasks or Nemotron 3.5 transcription. Check model-specific file limits; Chat Studio accepts text only.
Last reviewed on November 6, 2026
Use an audio upload for a supported video-reference task or for transcription with Nemotron 3.5 Transcribe. Check the exact model and operation before uploading; Chat Studio still accepts text only.
Choose the Right Use for the Recording
| What you need | Where to start | What the file does |
|---|---|---|
| A portrait animated with recorded speech | Kling AI Avatar Standard, Kling AI Avatar Pro, or OmniHuman 1.5 | One image and one audio file produce a video; this is not a cleaned audio export |
| An existing video matched to supplied audio | Volcengine Video Lip Sync | One video and one audio file are inputs to a new video result |
| Audio as a reference for video generation | Seedance 2.0 or Seedance 2.0 Fast, in reference mode | The recording guides a video request; it is not a promise to repair or reproduce the source exactly |
| New speech, scripted dialogue, sounds, or lyrics | Audio Studio and audio workflows | Use text and the chosen model's controls. These operations do not clean or transcribe an old recording. |
| Noise cleanup or speech isolation from an existing recording | See the current audio-cleanup explanation | ElevenLabs Audio Isolation is paused; its documented input is not an available cleanup service |
| A recording to transcribe | Nemotron 3.5 Transcribe | Upload one MP3 or WAV; the result is plain text, without word timestamps, translation or video input |
| A portrait animated with audio | OmniHuman 2.0 | One image and one audio file produce a video; check its separate limits below |
Text-to-speech, dialogue, and Suno Sounds generate new audio. They do not repair an uploaded recording. Currently available Chat Studio models accept text only, so you cannot send a recording there for transcription or a listening review. These generators and video-reference workflows do not establish a general voice-cloning service.
Match the File to the Model
Use the selected form and reference-file requirements to check the input contract. These examples have separate model-specific limits. File sizes below use MiB (binary megabytes).
| Model or mode | Audio requirements | Other inputs and limits |
|---|---|---|
| Seedance 2.0 / 2.0 Fast reference mode | MP3 or WAV, up to 15 MiB per file; 2–15 seconds per file | Up to 3 audio references, with a combined audio duration of no more than 15 seconds; check separate image and video requirements |
| Kling AI Avatar Standard / Pro | MP3, M4A, WAV, AAC, or OGG; Rivya's upload limit is 95 MiB; duration must be greater than 0 and no more than 300 seconds | Exactly one JPEG or PNG image, up to 10 MiB, and one audio file |
| OmniHuman 1.5 | MP3, M4A, WAV, AAC, or OGG, up to 10 MiB; duration must be greater than 0 and strictly less than 60 seconds | Exactly one JPEG, PNG, or WebP image up to 10 MiB, and one audio file |
| Volcengine Video Lip Sync | MP3, M4A, WAV, AAC, or OGG, up to 10 MiB, with a readable positive duration | Exactly one supported video and one audio file; video properties and Lite/Basic settings have their own requirements |
| Nemotron 3.5 Transcribe | One MP3 or WAV, up to 95 MiB, with a duration greater than 0 and up to 3,600 seconds | No prompt is needed; the result is text |
| OmniHuman 2.0 | One MP3 or WAV under 15 MiB, 4–300 seconds | One JPEG, PNG or WebP under 30 MiB; sides 300–6,000 pixels, ratio 0.4–2.5; optional prompt up to 300 characters |
Audio clips can each pass Seedance 2.0's per-file duration check and still exceed its 15-second combined limit. Keep the passages needed for the task, shorten the excerpts, and re-export before uploading. Other Seedance versions and models have their own requirements.
Play the exported copy locally before uploading it. Renaming an extension does not convert a file, and a playable recording can still exceed a model's size or duration limit. Use browser and file troubleshooting if playback or upload confirmation fails. For OmniHuman, check the model page's prompt-language and length requirements as well; those apply to the prompt, not the recording's language.
Prepare a Copy You Can Safely Send
Keep the original on your device. In a separate copy, trim unrelated conversation and private material, then listen for missing words, distortion, or accidental silence. Preserve the complete passage the chosen task needs, including its beginning and end.
Confirm your right to use the recording, the speaker's voice, and any accompanying image or video for the intended output. Follow the safe-upload guide before uploading. Uploading sends the file beyond your browser before you generate; removing the form attachment is not confirmation of deletion from other systems. Use only the material the request needs.
Upload and Send One Request
Sign in and open the chosen model and operation. Save unsent text, model settings, and source files locally if sign-in or a page change is needed; see login and draft protection.
Select the prepared audio and any required image or video. In Seedance 2.0 reference mode, use the audio-reference area. Wait for upload completion, then check the files, roles, counts, and durations where shown.
Review the selected mode, settings, and estimated credits, then submit once. Upload completion does not start generation. Changing the mode or model can reset settings or remove incompatible references, so check them again before sending.
Check the existing task in History if the response or connection is unclear. Do not immediately repeat the request. Follow queue and waiting guidance for an in-progress task and failed tasks and credit returns for a recorded failure.
Preparing or uploading a file does not guarantee acceptance, successful generation, or a particular quality. A second generation request can have a separate credit cost.
Review the Result for Its Actual Purpose
For avatar or lip-sync video, watch and listen to the whole result. Check for missing words, mouth timing, facial changes, unintended content, and problems at the beginning or end of the speech. For audio-guided video, check whether it follows the intended rhythm or sound direction; it may not preserve every detail of the recording.
Compare the result with the original before replacing or sharing anything. Save a usable output according to output downloads, and retain the original separately. An expired link or a preview that will not play does not establish that generation failed or that a credit return is due. Check the task record and actual file.
Resolve Rejections Without Repeating the Same Mistake
A missing audio field can mean the selected operation does not accept recordings. A paused cleanup model cannot be made available by paying for a plan or uploading a different file. Choose another available operation only if it produces the kind of output you need.
If a file is rejected, read the message and check its format, size, duration, companion files, and the selected mode's file-count or combined-duration limits. Fix the issue in a separate copy and play it locally before uploading again. If a task was already submitted, check its record before starting a replacement.
If you still need help, use Contact. Include the model, mode, file format, size, duration, visible error, approximate time, and task identifier if one exists. Upload problems can occur before a task identifier is created. Do not send passwords, API keys, private access links, or the sensitive recording itself in the initial message.
