Best Speech-to-Text APIs for Developers: 11 Services Compared

Key Takeaways

• Eleven services are compared without claiming an independently measured accuracy or latency ranking.

• whisper-1 is a deprecated batch model listed at $0.006/minute until its scheduled February 26, 2027 shutdown; it is not a live streaming model.

• Trial credits, browser allowances, recurring free quotas, and paid API usage are different offers.

• Language counts are model- and feature-specific and may include regional variants.

• A shared API key does not establish identical streaming protocols or response schemas across all providers.

How These Options Were Compared

This comparison uses official model documentation, model cards, published pricing, and lifecycle information checked on November 6, 2026 (UTC). The shortlist is organized by workload and deployment requirements, not an independently measured performance ranking. Published limits describe the cited model or checkpoint; hosting platforms may impose different limits. No private accuracy, latency, or successful-task measurements are claimed.

Speech-to-Text API Comparison

Signup credits and advertised maximum free hours are not recurring monthly quotas unless explicitly stated. Free offers and support conditions were checked November 6, 2026 (UTC). Languages, timestamps, diarization, and extra analysis features need model-level validation.

The 11 Speech-to-Text APIs, by Developer Requirement

1. Deepgram — Real-time voice and conversational transcription

Deepgram separates Nova transcription models from Flux conversational speech recognition. Flux adds turn-oriented behavior relevant to voice agents, while Nova is a separate model choice. The pricing page shows model-, mode-, and plan-specific minute rates, including promotional and regular figures. The $200 entry allowance is a trial credit balance, not a monthly quota. No independent fastest-latency or lowest-WER claim is made here. (Source: Deepgram transcription pricing).

2. AssemblyAI — Meeting notes and speech-understanding pipelines

AssemblyAI combines transcription with optional speech-understanding capabilities. Model choice determines language coverage; Universal-2’s 99-language coverage should not be transferred to every streaming model. Its free plan advertises up to 185 pre-recorded hours and up to 333 streaming hours, which are maximum offers rather than a stated recurring monthly renewal. Diarization, summaries, redaction, and other features need their own compatibility and billing checks. (Source: AssemblyAI transcription pricing).

3. OpenAI whisper-1 — Legacy batch transcription and migration

The hosted whisper-1 model is listed at $0.006 per minute but is deprecated, with shutdown scheduled for February 26, 2027. OpenAI recommends gpt-live-transcribe or gpt-transcribe as replacements. The same shutdown notice covers GPT-4o Transcribe models; self-hosted Whisper checkpoints are a separate deployment choice. See OpenAI’s transcription deprecation notice. The API guide lists 98 supported languages, with accuracy varying by language. A benchmark for Whisper Large v3 does not establish the hosted whisper-1 endpoint’s WER. Newer transcription or realtime APIs must not be described as capabilities of whisper-1 itself. (Source: OpenAI whisper-1 transcription specifications).

4. Google Cloud Speech-to-Text — Google Cloud speech pipelines

Google offers different recognition versions, models, regions, and processing modes. The published V1 allowance does not automatically apply to V2. The Standard V2 starting volume tier lists $0.016 per minute, while dynamic batch and discounted consumption modes have different rates. Its service-level 125 languages and variants are not a guarantee that every model supports each locale or feature. (Source: Google Cloud Speech-to-Text pricing).

5. Amazon Transcribe — AWS applications and contact centers

Amazon Transcribe integrates with AWS storage, identity, and contact-center services. General transcription, Medical, and specialized features have different scope and prices. The free offer starts with the first request and depends on AWS eligibility and terms. HIPAA eligibility does not establish that a customer application is compliant: service configuration, agreements, access controls, and the intended data flow still matter. (Source: Amazon Transcribe pricing).

6. Microsoft Azure Speech — Microsoft-stack enterprise speech

Azure’s free F0 allowance covers real-time Standard and Custom recognition, not batch transcription. Locale, region, custom-model support, and network configuration affect implementation. The language matrix includes regional variants and feature-specific limitations. Azure documents speech customization options; any accuracy gain on specialist vocabulary must be measured using representative audio and reference transcripts. (Source: Azure Speech pricing).

7. ElevenLabs Scribe — Combined speech generation and recognition products

The current API pricing page advertises 90+ languages for Scribe v2 and Scribe v2 Realtime; confirm the selected model’s language list and feature support. It lists $0.22 per hour for Scribe v2 batch transcription and $0.39 per hour for Scribe v2 Realtime. Keyterm prompting and transcript editing have separate charges. Its API pricing table includes a Free / Pay as you go column, so evaluate API-specific credits and billing terms rather than substituting a browser-only allowance or assuming unlimited free transcription. (Source: ElevenLabs Scribe API pricing).

8. Gladia — Multilingual and mixed-language transcription

Gladia’s self-serve async Solaria-3 rate starts at €0.16 per hour, while the real-time offering starts at €0.68 per hour. Add-ons and monthly volume change the bill. Its €50 credit offer is a trial balance, not recurring free hours. Language coverage, diarization, translation, and retention settings depend on engine and plan. A single response format does not make all audio-intelligence features free. (Source: Gladia transcription pricing).

9. Speechmatics — Accent, multilingual, and deployment-sensitive evaluation

The current price page lists Melia 1 pre-recorded transcription at $0.12 per hour and real-time transcription at $0.20 per hour, with 56 languages. Enhanced and medical models have separate rates and behavior. Deployment and residency requirements require confirmation for the chosen commercial offering; vendor location does not establish where every request is processed. Use accented, noisy, and domain-specific audio to compare accuracy instead of assuming the published language count predicts WER. (Source: Speechmatics transcription pricing).

10. Rev AI — Recorded audio with optional human-transcription workflows

Rev AI advertises 57+ languages across its API offerings, so describing it as an English-only platform is inaccurate. Reverb ASR, foreign-language transcription, streaming, and human transcription are distinct products with different rates and outputs. Human review is not assumed to be an automatic escalation inside every machine-transcription request. Trial credits and language support should be tied to the selected product. (Source: Rev AI transcription pricing).

11. IBM Watson Speech to Text — Existing IBM and domain-customization deployments

IBM’s Lite allowance provides a defined recurring evaluation path. The service supports different model generations and locale variants, so a count from one telephony table should not be treated as the entire service’s language limit. Customization and enterprise deployment choices make it relevant to existing IBM environments, but on-premises availability, commercial terms, and supported features require product-specific confirmation. (Source: IBM Watson Speech to Text overview).

How to Measure Transcription Accuracy

Word error rate is (substitutions + deletions + insertions) divided by the number of words in a reference transcript. A comparison needs the same audio, reference policy, punctuation/number normalization, and handling of overlapping speakers. It should not combine results from different datasets into a universal winner.

Include accents, background noise, quiet speech, specialist terms, names, code-switching, and overlapping speakers. Evaluate speaker attribution and timestamp accuracy separately from word error rate. Vendor-published results are useful evidence about a specific setup, not an independent benchmark for every production workload.

Batch Versus Streaming: Different Integration Contracts

Batch transcription typically uploads a file or supplies an accessible URL, then returns a result directly or through polling/webhooks. Streaming sends audio over a supported realtime connection and receives interim and final transcript events. The protocol may be WebSocket or gRPC; it is not identical across all providers.

Interim results can be revised. Voice-agent applications also need endpointing or turn detection, audio encoding validation, interruption handling, and reconnect behavior. A stream of a completed recording is not necessarily the same product as live microphone transcription. For voice agents that also need spoken responses, compare text-to-speech APIs separately from transcription.

The hosted whisper-1 API is batch-oriented and scheduled to shut down on February 26, 2027. OpenAI recommends gpt-live-transcribe or gpt-transcribe for migration; their capabilities and request contracts must be checked separately. Google’s recognition protocols and cloud SDKs should not be described as interchangeable with a Deepgram WebSocket client.

Pricing: Compare the Full Audio Workload

Normalize rates before comparing: $0.006/minute equals $0.36/hour. Preserve the quoted currency; Gladia’s euro rates are not US-dollar rates. Minimum billable duration, channels, model choice, volume commitments, diarization, redaction, translation, and storage can affect the bill. Subscription browser seats are not API inference pricing. A trial that lasts until credits run out is not an ongoing free tier. Self-hosting a Whisper checkpoint removes an external minute-based tariff but introduces compute, utilization, maintenance, and availability costs.

Speech-to-Text Model Access on GPT Proto

GPT Proto lists GPT-4o Transcribe for audio-to-text. OpenAI has scheduled the upstream model’s shutdown for February 26, 2027; gateway availability, support deadlines, and migration options must be confirmed separately with GPT Proto. Do not assume that GPT Proto exposes Deepgram, AssemblyAI, or every service in this comparison.

GPT Proto provides access to supported models through one account, one API key, and a shared prepaid balance. There is no required monthly subscription, and unused credits do not expire under the current pricing terms. The pricing page currently starts credit top-ups at $15; this adds usage balance rather than buying unlimited inference. Model-specific rates, cache charges, media units, and promotions remain separate. See current GPT Proto pricing.

GPT Proto documents endpoint-specific availability and compatibility: a model publisher’s native API features are not automatically available through every GPT Proto route. Applications remain responsible for retrieval, tool execution, permissions, output validation, and error handling.

Use the selected route’s documented upload format, model ID, output schema, and billing unit. Do not assume that a batch route provides streaming, diarization, or the upstream provider’s full locale matrix.