AirCaps A5S Async v1: Speaker-Attributed Transcription Benchmarks, Price & SRT

Why AirCaps A5S Async v1 Beats NVIDIA Nemotron 3.5 on Overlapping Speech

Standard Word Error Rate (WER) only checks whether the transcribed words match the reference text. Multi-speaker benchmarks use cpWER, which scores a word as an error both when it is misheard and when it is assigned to the wrong speaker label. Across VoiceArena's 139 natural English conversations (115 in-person room meetings and 24 online calls totaling 272,690 reference words), AirCaps A5S Async v1 finished first at 23.89% cpWER, well ahead of second-place NVIDIA Nemotron 3.5 Transcribe at 32.43% and the previous leaderboard holder, ElevenLabs Scribe v2, at 35.73%.

What stood out to me when inspecting the VoiceArena error breakdown is how AirCaps beats Nemotron 3.5 Transcribe. Total cpWER is the sum of three error types: deletions (missed words), substitutions (misheard words or words handed to the wrong speaker), and insertions (hallucinated extra words).

Look at the deletions column first: Nemotron 3.5 Transcribe actually misses slightly fewer words than AirCaps (8.32% vs. 9.29%). Unlike older APIs that simply drop audio whenever two voices collide—on public English meeting sets, AssemblyAI deletes 25.8% of reference words and ElevenLabs deletes 18.8%, compared to 8.0% for AirCaps—both NVIDIA and AirCaps hear the overlapping speech.

The gap opens up in the Substitutions column. Nemotron 3.5 posts an 18.43% substitution rate compared to 10.88% for AirCaps—meaning Nemotron makes 69% more substitution and speaker-attribution errors. When three people argue across a table, Nemotron captures the syllables but frequently tags the interjection to whichever speaker had the floor a second earlier. AirCaps keeps the speaker boundaries intact and cuts hallucinated word insertions down to 3.72% (versus 5.68% on Nemotron and over 20% on smaller models like Mistral's Voxtral 3).

That speaker-separation advantage scales as recordings get messier. VoiceArena divides its test suite into 10 subgroups by speaker count (2, 3, 4, and 5–8 speakers), acoustic environment (online vs. in-person), and overlap density (<15%, 15–30%, and 30%+). A5S Async v1 ranks #1 in all 10 subgroups. In 3-speaker conversations, Nemotron makes 48% more errors than AirCaps; in segments where 50% or more of the dialogue overlaps, the next-best model makes 61% more errors.

Across the five standard public meeting corpora—three in English (NOTSOFAR-1, AMI, CHiME-6) and two in Mandarin (AISHELL-4, AliMeeting) covering 226 recordings and 59 hours—A5S Async v1 achieves a macro average cpWER of 21.51%, beating Nemotron 3.5 Transcribe (30.40%) and AssemblyAI Universal-3.5 Pro (32.00%). Its Chinese performance is particularly strong, hitting 14.39% cpCER on AliMeeting against Nemotron's 25.72%, and 18.43% on AISHELL-4 against Nemotron's 31.11%.

---

$0.18/hr Pricing vs. Deepgram, AssemblyAI, and ElevenLabs

Most speech API providers advertise a low base rate per minute and then tack on per-hour surcharges once you turn on speaker diarization, keyterm boosting, or smart number formatting. AirCaps prices A5S Async v1 at a flat $0.18 per audio hour, billed per millisecond of audio, with speaker labels, word-level timestamps, keyterm biasing, smart formatting, and SRT/VTT subtitle export all included in that single rate.

(Note: Microsoft's $0.10/hr rate is a temporary promotional price running through December 31, 2026, after which it reverts to its $0.54/hr list price.)

When you compare like-for-like bills with speaker labels enabled, AirCaps is 33% cheaper than ElevenLabs ($0.27/hr), 36% cheaper than AssemblyAI ($0.28/hr), and less than half the price of Deepgram Nova-3 ($0.39/hr). Deepgram is still the fastest API on the market, returning a one-hour recording in a median of 6.3 seconds compared to AirCaps' 9.8 seconds (AirCaps runs at 340x real-time on a single H100 GPU). However, Deepgram Nova-3 struggles heavily on multi-speaker overlap, posting a 66.69% macro cpWER across the five public meeting benchmarks and 71.89% on VoiceArena.

When I ran the unit economics for a podcast clipping and captioning workflow, the $0.18/hr rate made the math straightforward. Transcribing a 10-minute two-person podcast episode costs exactly $0.03. If you run a micro-SaaS that charges creators $9.90 for 1,000 credits and deducts 1 credit per minute of video ($0.099 revenue for that 10-minute clip), your gross margin sits right around 70% before Stripe fees. Even if you take the finished speaker-attributed transcript and feed it into a lightweight text model ($0.10 per 1M input tokens, $0.50 per 1M output tokens) to extract meeting decisions, action owners, and deadlines, your total API cost for a one-hour call stays under $0.20.

If your workflow requires live, sub-second captions during an active Zoom call rather than post-call file processing, A5S Async v1 will not fit—its streaming sibling (A5S Streaming v1) is still in closed early access. For live WebSocket streams, look at our breakdown of ElevenLabs Scribe v3 and Scribe v3 Live, which handles 150ms real-time captioning, or self-host NVIDIA Nemotron 3.5 Transcribe if you have local GPU VRAM and need 40 languages offline.

---

Supported Languages, API Setup, and Direct SRT Subtitle Export

AirCaps A5S Async v1 accepts any audio or video format supported by ffmpeg from 0.1 seconds up to 3 hours in length. You can pass a public or pre-signed URL up to 5 GB (audio_url), or upload a local binary file up to 1 GB via POST /v1/files (which returns a temporary file_id stored for 24 hours). Multi-channel files are automatically mixed down to mono from the first audio stream.

Before you send a request to POST https://api.aircaps.com/v1/transcripts, there are four parameter rules in the AirCaps API documentation that will save you from failed jobs or garbled output:

• No Automatic Language Detection (language_code): A5S Async v1 supports exactly 9 languages: English (en, plus regional tags like en-US, en-GB, en-AU, en-IN, en-SG), Spanish (es, es-MX), French (fr), German (de), Portuguese (pt, pt-BR), Hindi (hi), Japanese (ja), Korean (ko), and Chinese (zh). There is no auto-detect for recorded audio. If you omit language_code, the API defaults to "en". Sending a Mandarin or Spanish recording without specifying "language_code": "zh" or "es" forces the model to map foreign speech into English words.
• The keyterms HTTP 400 Trap on Non-English Audio: You can pass an array of up to 60 phrases in keyterms (each phrase up to 6 words and under 50 characters) to bias recognition toward company names, guest names, or technical jargon. However, keyterms only works when language_code is English. If you include a keyterms array on a Chinese (zh) or Japanese (ja) request, the API immediately rejects the call with a 400 invalid_request error.
• Cross-Language smart_format: Setting "smart_format": true automatically converts spoken numbers, currency amounts, percentages, dates, and times into digits (turning "twenty five dollars" into "$25"). Unlike keyterms, smart_format works across all 9 supported languages and defaults to false.
• Tuning for 7+ Speakers (large_meeting): For standard interviews and podcasts with 2 to 6 speakers, leave large_meeting at its default (false). If you are transcribing an all-hands meeting, courtroom recording, or panel with 7 to 9+ speakers, set "large_meeting": true so the diarization pass allocates additional speaker clusters instead of merging infrequent voices.

Step 1: Create the Transcription Job

Sign up at the AirCaps console to claim the 5 free hours (there is currently no portal to add funds, and your quota resets each billing cycle; an exhausted quota returns 402 insufficient_balance). Submit your audio file with curl:

curl https://api.aircaps.com/v1/transcripts \
  -H "Authorization: Bearer $AIRCAPS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
        "model": "a5s-async-v1",
        "audio_url": "https://example.com/roundtable-episode.mp3",
        "language_code": "en-US",
        "smart_format": true,
        "large_meeting": false,
        "keyterms": ["AirCaps", "Nemotron", "cpWER"]
      }'

The endpoint returns a job id and sets the status to processing (or queued). You can poll GET /v1/transcripts/{id} or pass a webhook_url in the POST body to receive a callback as soon as the job finishes.

Step 2: Export Speaker-Labeled SRT or VTT Subtitles

Once the transcript status is completed, you do not need to write a custom script to convert word timestamps into subtitle blocks. Call the dedicated /subtitles endpoint on the same transcript ID:

curl "https://api.aircaps.com/v1/transcripts/{id}/subtitles?format=srt&chars_per_caption=42" \
  -H "Authorization: Bearer $AIRCAPS_API_KEY" \
  -o episode-captions.srt

The format query parameter accepts srt (default) or vtt (if you are embedding captions in a browser <track> element, see our guide on SRT vs VTT). The chars_per_caption integer controls line length and must be set strictly between 10 and 200 (default is 42). When I export subtitles for 16:9 YouTube videos, 42 keeps captions clean on a single line; when cutting 9:16 vertical shorts for TikTok or Reels, I drop chars_per_caption to 26 so the speaker tag and text never wrap into three stacked lines.

The returned application/x-subrip file embeds the speaker label directly at the start of each turn:

1
00:00:00,480 --> 00:00:02,310
[speaker_0] Hey, good afternoon, everybody.

2
00:00:02,400 --> 00:00:04,150
[speaker_1] Thanks for having me on the show.

If you regularly transcribe the same podcast hosts or team members, you can pair A5S Async v1 with AirCaps Speaker ID ($0.02 per audio hour). Once a person's voiceprint is enrolled, the /subtitles endpoint automatically replaces generic tags like [speaker_0] and [speaker_1] with their actual names (for example, [Alex] and [Sam]) right inside the exported SRT file.

---

Production Limits: The 5-Job Concurrency Cap, Data Privacy, and Open Benchmark Verification

Before wiring AirCaps A5S Async v1 into a live user-facing app, account for three operational constraints that don't show up on headline benchmark charts.

1. The 5 Concurrent Jobs Limit and FIFO Queue

By default, every AirCaps account is capped at 5 concurrent asynchronous jobs running at the same time (shared across transcription and standalone diarization requests). If a user uploads a batch of 25 meeting recordings at once, the first 5 process immediately (~10 seconds per hour of audio) and the remaining 20 enter a First-In, First-Out (queued) state.

The queue holds up to 2,000 pending jobs per account, and the job creation endpoint (POST /v1/transcripts) enforces a rate limit of 600 requests per minute. Exceeding either the 2,000-job queue or the 600 RPM creation cap returns HTTP 429 Too Many Requests with a Retry-After header. If you expect traffic spikes where dozens of users request short clips simultaneously, build a webhook listener rather than synchronous polling, or contact AirCaps to raise your concurrency ceiling before launch.

2. Default Model Training Policy and How to Opt Out

Out of the box, AirCaps states that customer audio and transcripts on standard accounts may be used to train and benchmark future models. Uploaded source audio files are cached for 24 hours (allowing you to re-run jobs on the same file_id) and permanently wiped from their servers within 24 to 48 hours, while generated transcripts remain stored for 30 days by default (deleting a transcript via the API immediately deletes the underlying audio file as well).

If you are transcribing confidential client calls, medical interviews, or unreleased episodes, upgrade to a paid balance and open the AirCaps console (Settings -> Data controls). From there, you can toggle off model training and shorten transcript retention down to 1 hour (or request an enterprise zero-data-retention agreement).

3. Verifying Accuracy on Your Own Audio with the Open Benchmark

Because a5s-async-v1 is a closed cloud model with no downloadable weights, you cannot inspect the architecture locally. However, AirCaps published their entire evaluation suite under an MIT license at github.com/aircaps/a5s-async-v1-benchmark, including the dataset preparation scripts, normalizers, cpWER scoring pipeline, and raw API outputs for every competing model.

Instead of trusting aggregate leaderboard numbers, take your noisiest 15-minute multi-speaker recording, run it through the 5 free hours on your account, and score the output against your own ground-truth transcript using their repo before switching your production pipeline over.