Hugging Face Inference Endpoints let you deploy a dedicated replica of an ASR model and pay for GPU or CPU time while that replica is initializing or running. On the Automatic Speech Recognition catalog (checked October 31, 2026), both whisper-large-v3 and whisper-large-v3-turbo list a catalog recipe around $0.8 per hour on 1× NVIDIA L4. Hugging Face does not publish a dollars-per-audio-minute rate for these dedicated Whisper endpoints.
The LecSync Realtime Speech API is a managed WebSocket for live meeting and lecture captions: preview and final text, translation between any two of 60 languages, a speaker and audio-aligned timestamps on every final segment, and running minutes. It costs $0.006 per audio minute ($0.36 per hour) with those features included.
Figures below were checked on October 31, 2026 against Hugging Face Inference Endpoints pricing, the ASR catalog, the whisper-large-v3 and whisper-large-v3-turbo deploy recipes, the transcription tutorial, and the LecSync developers page. Inference Providers pass through third-party rates without a clear public Whisper $/audio-minute on owned docs today, so this comparison sticks to dedicated Inference Endpoints compute pricing.
What is a Hugging Face Whisper Endpoint vs a realtime caption API?
A Hugging Face Whisper Inference Endpoint is a dedicated ASR deployment: you pick hardware (often 1× L4 for the catalog Whisper recipes), send audio files or HTTP transcription requests, and receive text. Hugging Face bills compute minutes while replicas run. A realtime caption API such as LecSync keeps one managed socket open for live bilingual captions, speakers and minutes while people are still talking, and bills accepted audio duration.
The short answer
If you want your own Whisper (or other open ASR) stack on Hugging Face hardware — custom models, scale-to-zero for bursty file jobs, and full control of the replica — Inference Endpoints fit. Idle scale-to-zero time is not billed as running compute; always-on 1× L4 is about $0.80 per wall-clock hour of GPU rent (~$576 per 24×30 month), which is compute rent, not an audio-minute invoice.
If you need live bilingual captions + speakers + digest on one flat minute bill, LecSync is the closer product. Whisper Endpoints are not a managed live-caption WebSocket with any-to-any translation, speakers and minutes packaged into $0.006 per audio minute.
How each path works
LecSync. Your server calls POST https://api.lecsync.com/v1/realtime/connect with the audio format, a language pair and the features you want, then streams 16 kHz mono PCM. Preview text arrives while someone is still talking; a final segment settles once the sentence confirms. Translations share the segment ID. With speakers: true, each final carries a speaker and startMs/endMs on the recording clock. With digest: true, minutes arrive section by section. The first word shows up about 0.28 seconds after speech starts; translation adds roughly 88 ms. Seven gateway regions (dal, jp3, kc, nj, ru2, ru3, uk) keep audio near the capture point; backlog is capped at 2 seconds.
Hugging Face Inference Endpoints (Whisper). From the Model Catalog you deploy openai/whisper-large-v3 or openai/whisper-large-v3-turbo with a verified recipe: NVIDIA L4, 1× GPU · 24 GB, ~$0.8/h in the UI checked today (us-east-1 shown on the recipe pages). The transcription tutorial walks through HTTP file upload to an /audio/transcriptions-style endpoint and optionally a second text-generation endpoint for post-hoc summaries. Catalog recipes expose Automatic Scale-to-Zero. Pricing docs state accounts are charged for compute while Endpoints are initializing and running; prices are shown hourly and billed per minute of that compute. Scale-to-zero replicas that are not running do not accrue running-compute charges.
Pricing: published compute vs audio minutes
| Path | Published figure (Oct 31, 2026) | What you actually buy |
|---|---|---|
| LecSync Realtime Speech API | $0.006 / audio minute all-in | Accepted audio duration; STT + MT + speakers + digest |
HF AWS nvidia-l4 ×1 | $0.8 / hour | GPU rent while replica runs (billed per compute minute) |
HF GCP nvidia-l4 ×1 | $0.7 / hour | Same metering on GCP L4 |
| Catalog whisper-large-v3 / turbo | ~$0.8 / hour (1× L4 recipe) | Same class as AWS L4 catalog default |
Always-on math (Calculated, compute rent): 1× AWS L4 at $0.80/hr → ~$19.20/day → ~$576/month if left up 24×30. That bill does not automatically equal “hours of audio transcribed.”
Estimated audio-minute conversions (assumptions stated): Hugging Face publishes no Whisper $/audio-minute for Inference Endpoints. Any $/audio-min figure is Estimated:
- Always-on 1:1 live coverage (one wall-clock hour of GPU for roughly one hour of live audio on that replica): Est. ~$0.80 per audio hour ≈ Est. ~$0.0133 per audio minute before concurrency — higher than LecSync’s $0.006, and still same-language ASR only.
- Batch faster-than-realtime with scale-to-zero (GPU up only while jobs run; Whisper often finishes audio faster than realtime): effective Est. $/audio-minute can fall well below $0.006, depending on throughput, queueing and cold starts. Label those scenarios Estimated; do not treat them as a published HF rate.
LecSync stays $0.006/min whether translation, speakers and digest are on or off → 1,000 audio hours = $360.
Nearby Whisper unit-price posts (different vendors, published $/audio-min): Groq Whisper, Cloudflare Workers AI Whisper, Together AI Whisper. Self-host ops without a cloud $/min: Whisper.cpp.
Side-by-side facts (checked October 31, 2026)
| LecSync Realtime Speech API | HF Inference Endpoints Whisper | |
|---|---|---|
| Primary job | Live meeting/lecture captions + MT | Dedicated Whisper ASR replica |
| Price model | $0.006/audio-min flat, all features | GPU/CPU $/hr while running; no public $/audio-min |
| Delivery | Managed WebSocket, preview→final | HTTP transcription / file upload to your endpoint |
| Live any-to-any text MT | 60 languages, same socket | Not packaged; DIY second model or offline translate task |
| Speakers on live stream | Every final (speakers: true) | Build yourself (tutorial notes Hub diarization models as a next step) |
| Live minutes / digest | digest: true | Separate LLM endpoint in the tutorial pattern |
| Languages | 60, auto-detected | Whisper-family multilingual recognition |
| Latency story | ~0.28 s first word; +~88 ms MT | Request / file inference; cold start if scaled to zero |
| Scale-to-zero | N/A (managed SaaS) | Catalog recipes support Automatic Scale-to-Zero |
| Regions | 7 gateways; 2 s backlog cap | Endpoint region you choose (recipe UI showed us-east-1) |
| Model control | Managed stack | Pick Whisper variant or other Hub ASR models |
What 1,000 hours cost (Estimated where labeled)
| Workload | LecSync | Hugging Face Whisper Endpoints |
|---|---|---|
| Same-language file / batch transcripts | Live API is not the cheapest file path | Estimated from GPU-hours × throughput; often cheap with scale-to-zero |
| Always-on replica for continuous coverage | $360 for 1,000 audio hours | Compute rent: e.g. one always-on L4 ≈ $576/mo (24×30) whether or not you fill it with audio |
| Translated live captions (arbitrary pair) | $360 | Whisper ASR + separate MT stack you operate |
| Speakers + timestamps during live session | Included | Not included as a managed live feature |
| Live minutes | Included | Extra generation endpoint / your own pipeline |
| Predictable all-in $/audio-min for captions + MT + speakers | Yes ($0.006) | No published audio-minute SKU |
For flat-vs-add-on vendor math, see what 1,000 hours cost once translation, speakers and summaries are added. For stitching STT + MT yourself, see built-in translation vs a DIY pipeline.
When Hugging Face Whisper Endpoints win
- Custom Whisper or other Hub ASR models on hardware you size, with catalog recipes for large-v3 / turbo.
- Bursty file transcription where scale-to-zero keeps the bill near zero between jobs.
- Batch jobs faster than realtime, so one GPU-hour clears many audio-hours (Estimated effective $/min can undercut managed caption APIs).
- Own-stack control — private endpoints, HF auth, VPC options, paired LLM endpoints for post-call notes.
- Same-language transcripts only, with no need for live any-to-any captions on one socket.
When LecSync fits better
- Translated live captions on one socket — any pair of 60 languages, each translation tied to its segment.
- Speakers and recording-aligned timestamps during the live stream — meeting client with speakers and timestamps.
- Live minutes while the session runs (
digest: true) — live digest vs post-call summarization. - Predictable $0.006/audio-minute billing with no GPU sizing, cold starts or Estimated throughput math.
- Classroom / mixed-language lecture feeds that must keep captions current as the speaker switches languages.
Category winners
| Category | Winner | Why |
|---|---|---|
| Custom open Whisper / Hub ASR on rented GPU | Hugging Face Endpoints | Catalog recipes + model choice |
| Bursty batch STT with scale-to-zero | Hugging Face Endpoints | Pay only while replicas run |
| Published $/audio-min for live captions | LecSync | Flat $0.006 all-in |
| Live any-to-any captions + speakers + digest | LecSync | One WebSocket, one minute bill |
| Always-on wall-clock GPU rent transparency | Hugging Face | Clear $0.7–$0.8/hr L4 list rates |
| Managed first-word latency for live captions | LecSync | Public ~0.28 s / +~88 ms MT |
FAQ
Does Hugging Face publish a Whisper price per audio minute on Inference Endpoints?
No. The pricing page lists hourly GPU/CPU rates billed per minute of compute while Endpoints initialize and run. Catalog Whisper recipes show about $0.8/hr on 1× L4. Any $/audio-minute number is Estimated from assumptions about throughput or always-on coverage.
How does $0.80 per GPU hour compare to LecSync’s $0.006 per audio minute?
They meter different things. LecSync charges accepted audio minutes. Hugging Face charges replica minutes. An always-on L4 at $0.80/hr is roughly Est. $0.0133 per wall-clock minute of GPU — higher than $0.006 if that minute maps 1:1 to audio, and still without live MT, speakers or digest. Batch scale-to-zero workloads can reverse that comparison on Estimated effective $/audio-min.
Can Inference Endpoints replace a live bilingual caption product?
You can stream audio into your own service in front of an endpoint, but the owned HF docs and Whisper recipes describe deployed ASR for transcription requests / files, plus optional separate text models for summaries. They do not ship LecSync-style any-to-any live MT, speaker labels and digest on one flat audio-minute bill.
Is scale-to-zero free?
Pricing docs charge for compute while Endpoints are initializing and running. Catalog recipes include Automatic Scale-to-Zero; when no replica is running, you are not paying that running-compute rate. Cold starts apply when traffic returns.
When should a live-caption product still use Hugging Face?
When the heavy work is file or batch Whisper under your control, or you already standardize on Hub models and Inference Endpoints. For a managed realtime speech API at $0.006/min with translation, speakers and digest, start from the LecSync developers page.
Final verdict
Hugging Face Inference Endpoints Whisper wins when you want dedicated open ASR compute — custom models, scale-to-zero batch jobs, and transparent GPU-hour rent around $0.80/hr on 1× L4. LecSync wins when the product is a managed realtime speech API at $0.006 per audio minute all-in, with live translation, speakers and digest on one connection.
Next step: open the LecSync developers page, create a connect session, and stream three seconds of PCM to see preview→final captions on one bill.