fal.ai·Speech to text

Wizper (Whisper v3 -- fal.ai edition)

/fal-ai/wizperasync$0.003 / compute second

[Experimental] Whisper v3 Large -- but optimized by our inference wizards. Same WER, double the performance! Returns hosted file URLs (download promptly; fal keeps them for a limited time). Billed per compute second.

Charged per compute second produced; a call that produces nothing is charged one unit.

Input

versionstring
Version of the model to use. All of the models are the Whisper large variant. (default "3")
audio_urlstringrequired
URL of the audio file to transcribe. Supported formats: mp3, mp4, mpeg, mpga, m4a, wav or webm.
merge_chunksboolean
Whether to merge consecutive chunks. When enabled, chunks are merged if their combined duration does not exceed max_segment_len. (default true)
taskstring
Task to perform on the audio file. Either transcribe or translate. (default "transcribe")
transcribe, translate
languageany
Language of the audio file. If translate is selected as the task, the audio will be translated to English, regardless of the language selected. If `None` is passed, the language will be automatically detected. This will also increase the inference time. (default "en")
chunk_levelstring
Level of the chunks to return. (default "segment")
max_segment_leninteger
Maximum speech segment duration in seconds before splitting. (default 29)

Field	Type	Required	Description
version	string	—	Version of the model to use. All of the models are the Whisper large variant. (default "3")
audio_url	string	yes	URL of the audio file to transcribe. Supported formats: mp3, mp4, mpeg, mpga, m4a, wav or webm.
merge_chunks	boolean	—	Whether to merge consecutive chunks. When enabled, chunks are merged if their combined duration does not exceed max_segment_len. (default true)
task	string (transcribe, translate)	—	Task to perform on the audio file. Either transcribe or translate. (default "transcribe")
language	any	—	Language of the audio file. If translate is selected as the task, the audio will be translated to English, regardless of the language selected. If `None` is passed, the language will be automatically detected. This will also increase the inference time. (default "en")
chunk_level	string	—	Level of the chunks to return. (default "segment")
max_segment_len	integer	—	Maximum speech segment duration in seconds before splitting. (default 29)

Run it

curl -s https://sooma.dev/v1/run \ -H "Authorization: Bearer $SOOMA_API_KEY" \ -H "Content-Type: application/json" \ -d '{"provider":"fal","endpoint":"/fal-ai/wizper","input":{}}'

Or ask your agent: it will call inspect for this schema and run it over MCP.