VOICE AI INFRASTRUCTURE

Fast, accurate speech APIs at half the usual price.

Fast, accurate, cost-effective speech APIs: proprietary and open-weight models behind one gateway, with speaker intelligence and LLM processing built in.

OpenAI-compatible endpoints  ·  Batch + streaming  ·  From $0.0012/min
standup.wav · 2 SPEAKERS · 8.2s
00:00.000:02.000:04.000:06.000:08.2
Okay,sotheQ3DATErollout.Priya,PERSONwhereareweonthemigration?WeshippedthebatchpipelineTuesday.DATELatencyisdownfortypercent,PERCENTandstreaminglandsearlynextsprint.DATE
SPEAKER 1 SPEAKER 2ENTITIES: DATE · PERSON · PERCENT
POWERING VOICE FEATURES AT
CallvaneNOTABLYhearsay labsDIALPOSTQuorumminuteworksLoopdeskVOCALISFieldnoteAttica CX
01 · CAPABILITIES

Everything after the microphone.

01+
Speech-to-text

Batch and streaming transcription with word-level timestamps.

POST /v1/audio/transcriptions
02+
Diarization

Who spoke, when, separated into speaker turns.

POST /v1/audio/diarizations
03+
Speaker ID

Attach turns to known, enrolled voices.

POST /v1/audio/identifications
04+
Understanding

Formatting, entities, sentiment, topics on the same job.

"processing": [ ... ]
05+
LLM processing

Summaries, extraction, classification on transcripts.

POST /v1/responses
06SOON
Realtime

Conversational sessions with interruption handling.

/v1/realtime
02 · DEVELOPER EXPERIENCE

Feels familiar on purpose.

Ultrafield speaks the API you already use: OpenAI-compatible audio endpoints, Bearer auth, idempotency, resumable SSE, signed webhooks. Point your client at a new base URL and send audio.

OpenAI-compatible audio endpoints
Bearer auth + idempotency keys
Resumable SSE streams
Signed webhooks
TRANSCRIBE.SH
$ curl https://api.ultrafield.ai/v1/audio/transcriptions \
    -H "Authorization: Bearer $ULTRAFIELD_API_KEY" \
    -F file=@support-call.wav \
    -F model=ultrafield-asr-1 \
    -F processing=["diarization","entities","summary"]

HTTP/1.1 200 OK
{
  "text": "Thanks for calling. Could I get the order number?…",
  "words": [ { "w": "Thanks", "start": 0.12, "end": 0.34 }, … ],
  "speakers": [ { "id": "S1", "turns": 11 }, { "id": "S2", "turns": 9 } ],
  "entities": [ { "type": "ORDER_ID", "span": "A-4417" } ],
  "summary": "Support call about a delayed order; replacement shipped.",
  "usage": { "audio_seconds": 512, "llm_tokens": 486 }
}
03 · MODELS

Choose the engine. Keep the API.

MODELTYPEMODESBEST FOR
Ultrafield ASR-1PROPRIETARYBATCH + STREAMINGMax accuracyPRO
Whisper large-v3-turboOPEN WEIGHTBATCH + STREAMINGLow latency
Whisper large-v3OPEN WEIGHTBATCHHigh accuracy
pyannote speaker processingOPEN WEIGHTBATCHDiarization + ID
GPT-OSS-120BHOSTED LLM·Transcript processing
OpenAI · AnthropicGATEWAY·SOON
04 · PIPELINES

One request. The whole pipeline.

Chain formatting, speakers, entities, sentiment, and LLM steps in a single request. Independent steps run in parallel; you get one job, one webhook, one structured result.

audio
transcribe
format
diarize
entities
summary
webhook · json
INDEPENDENT STEPS RUN IN PARALLEL · ONE JOB · ONE WEBHOOK · ONE STRUCTURED RESULT
05 · ECONOMICS

Predictable at volume.

Prepaid credits, drawn down per processed second. Pay for the stages a request actually runs, not seats, not bundles.

TYPICAL PROVIDER$0.26/HR
ULTRAFIELD$0.13/HR
STAGEPER MINUTEPER HOUR
Speech-to-text (batch)$0.0021/min$0.13/hr
Speech-to-text (streaming)$0.0031/min$0.19/hr
Diarization$0.0010/min$0.06/hr
Speaker ID$0.0008/min$0.05/hr
BILLED PER SECOND
06 · WHY ULTRAFIELD

Why Ultrafield.

01+
100+ languages

Transcribe across 100+ languages with automatic language detection.

02+
State-of-the-art accuracy

Competitive with the best ASR available. Evaluate on your own audio and pin model versions for reproducibility.

03+
Low latency

Fast response times on optimized, chained infrastructure. Transcribe an hour of audio in as little as 10 seconds.

04+
Zero data retention

Your data is never stored beyond ephemeral processing. We keep only aggregated usage stats for account management.

05+
No training on your data

We never train models on customer audio, transcripts, or outputs. Zero-data-retention options available.

06+
US or EU data residency

Choose US or EU data centers for Enterprise workloads.

07 · REALTIMEROADMAP

Next: conversation.

Realtime voice-agent APIs are in development: natural turn-taking, interruption handling, tool execution, on the same platform and the same event conventions.

/v1/realtimeSOON
NATURAL TURN-TAKINGINTERRUPTION HANDLINGTOOL EXECUTION
08 · FAQ

Questions, answered.

Which speech models can I use?

Ultrafield's own models alongside open-weight options including Whisper large-v3 and large-v3-turbo. Choose directly, or let the pipeline match a model to the workload.

Batch and real-time?

Yes. Uploaded files, audio URLs, and low-latency streams share one API surface: archives, live captions, analytics, conversational products.

What can I add beyond transcription?

Diarization, speaker identification, formatting, entity detection, sentiment, and LLM processing, composed in one request.

Can it identify who is speaking?

Diarization separates a recording into speaker turns; identification attaches turns to enrolled voices when your workflow needs it.

How does LLM processing work?

Chain a transcript into hosted models such as GPT-OSS-120B for extraction, classification, or summaries. A gateway to OpenAI and Anthropic models is on the roadmap.

What does it cost at scale?

Per-minute rates published on the pricing page, designed to stay low as volume grows. Volume rates for sustained workloads.

Get started

Give your product better ears.

Move from raw audio to accurate transcripts, speakers, entities, sentiment, and application-ready output through one pipeline.

Batch and streaming · Proprietary and open models