Fireworks turns an hour of audio into text in under 4 seconds, serving Whisper v3-large on its own stack with alignment and translation
Synopsis
Fireworks released a beta speech-to-text API for the Whisper v3-large models, built around speed: on a dedicated endpoint, the asynchronous endpoint transcribes or translates one hour of audio in 4 seconds, about 900 times real time. Previously such services left users waiting seconds to minutes, and the gain here comes from serving-stack changes rather than a new model. On quality, whisper-v3-large reports a 2.00% word error rate for greedy inference on Librispeech Clean, matching the original open-source model; translation, word- and sentence-level timestamp alignment, audio preprocessing and voice activity detection ship in the same API.
Interpretation
Fireworks launched a beta speech-to-text API supporting the Whisper v3-large models, selling transcription speed and bundling translation, word- and sentence-level timestamp alignment, audio preprocessing and voice activity detection. Earlier transcription services made users wait seconds to minutes for a piece of audio, especially in bulk processing of long files; Fireworks treats low latency as a precondition for audio products to be usable at all. On a dedicated endpoint with 1 Fireworks Audio Unit, the asynchronous speech-to-text API transcribes or translates one hour of audio in 4 seconds, or 900 times real time; Fireworks reports latency 2 to 400 times lower than other Whisper providers, including 23 times compared with OpenAI.
The speedup did not come at the cost of recognition quality, and transcripts stay consistent with the original open-source Whisper v3-large. Fireworks checked quality with word error rate while optimizing speed, rather than reporting throughput alone. whisper-v3-large reports a 2.00% word error rate for greedy inference on the Librispeech Clean dataset, matching the original open-sourced whisper-v3-large.
Beyond plain transcription, the API returns start and end timestamps for every word and sentence, serving uses such as video captions, training data for speech recognition models and podcast editing. Text-only transcription cannot support uses that need timing, so Fireworks made alignment an out-of-the-box option. Alignment supports three aligner models, Gentle, mms_fa and tdnn_ffn; Fireworks runs Gentle on GPU compute instead of CPU, reporting 10 seconds to align one hour of audio and up to 10 times more efficient compute for Gentle-based alignment with no accuracy impact.
Perspective
The demonstrated scope is the Whisper v3-large family on Fireworks' own dedicated endpoints: 4 seconds to transcribe one hour of audio asynchronously and 10 seconds to align one hour, tested on FLAC-encoded mono audio at a 16000 sample rate with 1 Fireworks Audio Unit. During the beta, serverless endpoints do not reach dedicated-deployment speeds, and both whisper-v3-large and whisper-v3-large-turbo are offered serverless.
The speed and cost multiples are Fireworks' own measurements, taken on a dedicated endpoint with 1 Fireworks Audio Unit; during the beta, serverless endpoints explicitly do not reach the same speeds, pricing is not yet enabled, and rate limits may be hit.