Hosted Speech Recognition APIs

aiOla sells Jargonic for transcription with specialized vocabulary, including healthcare, legal, finance and engineering terminology. Developers supply a custom dictionary without training a separate model. The documented family includes Jargonic v2 and v2 Flash; Flash trades some recognition accuracy for faster responses. The earlier v1 model remains listed in its documentation.

Alibaba Cloud operates Qwen ASR through Model Studio. Its Qwen3-ASR-Flash file interface accepts publicly accessible audio URLs or Base64 audio through an OpenAI compatible API. Developers can supply background text and entity glossaries as recognition context, and leave the language unspecified for multilingual recordings.

The asynchronous Filetrans model requires DashScope, while Flash supports both DashScope and OpenAI compatible requests. Beijing and Singapore have workspace-specific API domains. Temporary upload URLs expire after 48 hours and the upload credential API has a fixed 100 requests per second limit; Alibaba recommends stable storage for production audio, such as OSS.

Amazon Web Services operates Transcribe for streaming audio and recordings. Batch jobs read media from an S3 bucket. The service can label speakers and separate channels, customize recognition for vocabulary, and filter content. Language and feature support depend on the transcription mode.

Billing uses one-second increments without minimums, with rates set by AWS region. Personal information redaction and custom language models carry additional charges. Regional coverage also differs for standard Transcribe, Transcribe Medical and Call Analytics. Under AWS’s HIPAA agreement, protected health information requires encryption both at rest and in transit.

AssemblyAI has recorded and streaming APIs for products such as call analytics, meeting transcription and voice agents. Its asynchronous recording API covers 99 languages with add-ons for speaker labels and personal information redaction. Developers can also use synchronous HTTP transcription or a dictation endpoint with formatting instructions when they want cleaned up text returned in the same request. Python and TypeScript SDKs provide access to the API family through the same development tools.

Live recognition has its own model choices. Universal-Streaming costs $0.15 hourly, with English and six-language multilingual versions. Universal-3.6 Pro Realtime supports 32 languages at $0.45 hourly. Those are base recognition rates; speaker diarization and other options can add to the price.

The same platform sells summaries, sentiment and topics as speech understanding features. For budgeting, its rate card lists diarization as an extra charge, with different rates for recordings and streaming. Vocabulary prompting, medical mode and redaction also have their own model-specific prices, so a transcription job with several extras can cost more than the advertised base rate.

Microsoft operates Azure Speech, which has realtime and batch transcription alongside a fast transcription API. Developers can also add pronunciation assessment with reference text, for applications that give speakers feedback on spoken accuracy and fluency. Language coverage is documented separately from the individual transcription interfaces. Custom models can also adapt pronunciations and text formatting, including rewrite rules, inverse text normalization and profanity filtering settings.

Fast transcription returns a recording’s results through an HTTP request and accepts files under five hours and 500 MB. It takes an inline upload or audio from a public URL, and requires a Speech resource in a supported region. Developers can specify a known locale or request language identification.

For specialized vocabulary, Custom Speech trains adapted models using text or audio with reference transcripts. Its costs include model usage and endpoint hosting, with training charges for newer base models. Batch use does not require endpoint hosting. A custom endpoint can keep its selected model fixed until the buyer chooses to update it, subject to the model’s supported lifecycle.

Cartesia’s Ink family covers batch file transcription and live WebSocket recognition. The current Ink-2 model targets streaming applications and supports English, Spanish, French, Hindi and Japanese. The batch endpoint’s documented model is ink-whisper, so the two modes have different model choices.

Cartesia bills through monthly plans with shared credits. Free includes 20,000 credits, while Pro costs $5 a month for 100,000 credits and includes a commercial use license. Ink-2 consumes three credits per audio second. Speech recognition has its own concurrency pool, separate from speech generation and managed agents, with limits rising by plan.

Cohere Transcribe supports 14 languages and 25 MB files, including English, Arabic, Chinese, Japanese and several European languages. The hosted experimentation endpoint is free and subject to rate limits. Production buyers provision a dedicated managed Model Vault, with private cloud inference and pricing per instance hour; longer commitments can receive discounts.

The model expects one specified language at a time and has no explicit automatic language detection. It also omits timestamps and speaker labels. Those limits are relevant for subtitles and meeting products that need to align text with audio or attribute remarks to participants.

Corti’s stateless dictation WebSocket API returns transcripts and detected commands. Clients authenticate within a tenant and select a US or EU environment. A connection lasts up to four hours. Corti provides separate recording and ambient endpoints, so file transcription and ambient interaction capture use different integrations from live dictation.

Deepgram handles files and live audio for developers building transcription into their applications. Its Nova models cover more than 45 languages, with speaker labels, automatic language detection, text formatting and vocabulary prompts available across the family. The CLI can take an uploaded file, an audio URL, microphone input or piped audio, and export SRT or WebVTT for caption workflows. Its tooling also exposes summaries, topics and sentiment, along with structured JSON, YAML, CSV and plain text output.

There are different models for different jobs. Flux targets voice agents, with turn detection and interruption handling built into recognition. Nova is the broader transcription family. Deepgram publishes separate streaming and recording rates, so the model and input mode both affect the bill.

The entry plan includes $200 in free credit, followed by usage billing without a minimum commitment. The Growth plan starts with $4,000 in annual credits. Published pay as you go limits include 50 concurrent REST requests and up to 150 WebSocket requests, while Whisper Cloud has a separate limit of five.

DeepL’s Voice API creates live translation sessions over WebSocket, with source transcription and translated output for multilingual applications. Developers first request a session and receive the streaming URL and access token. DeepL’s language matrix identifies recognition supplied by external partners; those source languages must be specified explicitly because automatic detection is unavailable for them.

Speech to text translation requires API Growth or Enterprise. The free Developer plan excludes it. Growth combines a fixed subscription with included translation hours and usage charges above the allowance, and has a 300-hour monthly speech limit. Enterprise uses custom commitments for larger workloads.

ElevenLabs sells Scribe for recorded and live transcription. The current family includes Scribe v2, Scribe v2 Realtime and a medical variant. The file model covers more than 90 languages, returns word timestamps and can label up to 32 speakers. It also tags audio events and detects the language, giving media and conversation products more than a block of transcript text. The medical version uses the same API and supports the same languages and features as Scribe v2, with recognition adapted for clinical audio.

For domain vocabulary, file transcription accepts up to 1,000 keyterms. Other options include entity detection and editing transcripts with instructions. Live audio uses a WebSocket interface with its own realtime model.

The API price page lists Scribe v2 at $0.22 hourly and Scribe v2 Realtime at $0.39 per audio hour. Vocabulary prompting and transcript editing have separate hourly charges. Scribe v2 Medical has the same base rate as the general file model. Healthcare buyers must arrange a BAA through sales before using it in integrations that require HIPAA compliance.

Fireworks AI sells streaming and recorded audio transcription for enterprise voice applications. Its speech product includes speaker diarization and audio preprocessing, with use cases centered on live voice workflows. The platform also advertises language-aware recognition for production. Developers can start from the product page’s trial flow or discuss enterprise requirements with sales.

OVH Groupe owns Gladia following its acquisition, completed in July 2026. Gladia provides live and asynchronous transcription of audio and video, with tools for speaker labeling, custom vocabulary, subtitles and translation. Its documented integrations include LiveKit, Pipecat and Twilio, covering voice agents and telephony applications.

The commercial service uses prepaid credits without subscriptions, with automatic volume discounts and €50 in introductory credit. Beyond recognition, developers can request summaries, sentiment and entity extraction, or pass audio into an LLM workflow. The documentation has separate controls for concurrency, supported file duration and data retention.

Google Cloud operates Speech-to-Text for developer applications, with separate methods for short recordings, long files and live audio. The documentation includes language and model availability, plus automatic language detection and word confidence options. Buyers choose a service version as well as the model and processing region. The platform also accepts phrase sets and custom classes to guide recognition toward application vocabulary, with quotas on phrase counts and character lengths.

For V2 standard recognition, the first 500,000 monthly minutes cost $0.016 per minute, with lower rates in higher volume bands. Billing measures successfully processed audio in increments of one second. A processed request can incur a charge even when it returns an empty transcript.

The request limits shape implementation. Synchronous recognition accepts one minute or 10 MB, whichever comes first. Streaming sessions last up to five minutes before the application needs to continue through another session. Batch jobs read files from Cloud Storage, with eight hours per file. Multiple-language recognition is limited to the global, US and EU endpoints, so a regional deployment needs its own capability review.

Groq hosts Whisper Large V3 and Turbo behind OpenAI compatible file transcription endpoints. Both support multilingual recognition. Large V3 also supports translating speech into English, while Turbo is limited to transcription. Developers can request structured JSON output and timing information for text aligned to a recording.

The published audio rates are Turbo at $0.04 hourly and Large V3 at $0.111 hourly. The API has separate transcription and translation routes, so applications needing English output from foreign-language recordings must use the translation-capable model and endpoint. This product’s documented interface takes files for processing.

IBM operates Watson Speech to Text as a provisioned IBM Cloud service. Developers create an instance, obtain its credentials and send audio to the recognition endpoint. The setup guide demonstrates FLAC transcription through REST and recommends IAM tokens for production authentication. Language model selection is part of the recognition request.

Mistral AI’s current hosted lineup includes Voxtral Mini and Small Transcribe 3 for recorded files, plus Mini Transcribe Realtime 3 for live speech. Both modes support speaker labels and vocabulary context. File results can arrive in pieces, while live sessions return recognition events as audio arrives.

The feature limits differ by mode. Offline transcription has segment timestamps without word timestamps; realtime segments have speaker labels but no start or end timestamps. Realtime supports turn detection, but diarization needs a separate session because the service rejects a session that enables both. Mistral publishes a migration guide for applications moving from Voxtral 2.

CNTXT AI operates Munsit, an Arabic voice API with file uploads and live transcription. Separate endpoints add speaker diarization, sentiment analysis and meeting minutes. Developers get free credits without a card when signing up, and one API key covers the transcription and other voice endpoints. The documentation also lists LiveKit and Pipecat integrations.

OpenAI sells file transcription through gpt-transcribe. Developers send an audio file and receive text plus detected languages as JSON. The API accepts files up to 25 MB, including MP3, MP4, M4A, WAV and WebM. Larger recordings need compression or splitting into smaller files before submission. Applications receive detected language codes in JSON; when the model cannot predict a language reliably, that list is empty. For multilingual audio, developers can supply several language hints in the same request.

The current file model takes context and vocabulary hints, including a prompt, keywords and language hints. It can also stream the transcript response from a completed recording. Audio arriving continuously from a microphone or call uses the separate Realtime transcription interface.

The published rate for gpt-transcribe is $0.0045 per audio minute. Live transcription models have a $0.017 minute rate. Existing integrations also need a migration plan: Whisper and GPT-4o transcription retire on February 26, 2027, along with the legacy audio translation endpoint. The documentation retains compatibility examples for those older models until shutdown.

Rev operates Rev AI, with asynchronous and streaming recognition for developers transcribing recordings or live audio. Its API family also includes custom vocabulary and forced alignment, plus language identification, topic extraction and sentiment analysis.

The pay as you go rate card lists English Reverb at $0.20 hourly and foreign-language Reverb at $0.30 per hour. Both round usage to the nearest second with a 15-second minimum per job. New accounts get credits equivalent to five hours of Reverb recognition. Rev also sells human transcription through the API at $1.99 per minute, while the analysis services have separate usage rates.

Sarvam AI’s transcription API focuses on Indian languages and English. Its default Saaras v4 model supports 22 Indic languages alongside Indian and global English. Developers have REST for short interactions, batch for longer recordings and a streaming interface. The file API accepts common audio and video formats, including WAV, MP3, FLAC and MP4.

Speaker diarization is available through batch only, with separate pricing from ordinary transcription. The synchronous endpoint merges multiple audio channels into one. Raw PCM input requires 16 kHz audio and an explicit codec parameter; other common formats can have their codec detected automatically.

Smallest AI’s Pulse handles live and recorded multilingual audio, with speaker diarization, redaction and word timestamps. Its model card lists 21 streaming and 12 recording languages, plus regional language groupings for mixed-language input. Coverage depends on both the mode and deployment region.

The Standard plan lists streaming at $0.004 per minute and batch at $0.003, with 100 concurrent streams and 25 batch requests per minute. East Asian streaming codes are restricted to the US region. South Indian streaming languages are in beta in India, and European streaming languages are also marked beta.

Soniox has live and asynchronous transcription across more than 60 languages. Speakers can mix languages within one conversation. Developers use REST and WebSocket interfaces, with project-specific API keys managed in the Soniox console. The same platform also sells speech translation and speech generation.

Recognition is billed by token usage, with published equivalents of about $0.10 per hour for files and $0.12 for streaming. Custom context and generated text affect the token bill. Speaker labeling, automatic language identification and formatting are included with recognition. Soniox and Tencent Cloud have announced a technology partnership combining recognition with realtime infrastructure.

Speechmatics has managed cloud transcription APIs for recordings, live audio and voice agents. Its model choices include general transcription, medical speech and an agent model with built-in turn detection. Melia 1 handles mixed-language recordings across 56 languages without requiring a language choice before submission.

For recordings, Melia 1 costs $0.12 hourly, while Enhanced is $0.38 per hour and requires a selected language. Oak 1 Medical is $0.15 per recorded hour. New accounts receive $100 in credits without a card. Speechmatics also offers containers and on-premises deployments for buyers that need to run the service in their own environment.

Tencent operates Tencent Cloud ASR for recordings and realtime streams. Recording jobs accept audio URLs or local files and return results through callbacks or polling. URL submissions allow five hours and 1 GB, while local uploads are limited to 5 MB. The service retains results for 24 hours; speaker separation depends on the selected engine.