Sources
Every number on this site comes from one of these 33 sources, read on the date shown. Vendor reports come last and are marked. A vendor runs these tests on its own models and sometimes on competitors, so they are not independent.
Batch and streaming leaderboard data from Artificial Analysis.
Artificial Analysis STT streaming
Independent benchmark by Artificial Analysis. AA-WER Streaming weights AA-AgentTalk (50%), VoxPopuli (25%), and Earnings22 (25%). WER is from the final transcript; the first-partial WER is kept as a separate value. Final latency is the mean time from end of speech to the final transcript. Price is AA's normalized USD per 1,000 minutes. Each point is one model on one tested provider and endpoint. Methodology: https://artificialanalysis.ai/methodology/speech-to-text.
Artificial Analysis STT non-streaming
Independent benchmark by Artificial Analysis. AA-WER v2 is the audio-duration-weighted WER over about 8 hours of English audio: AA-AgentTalk (50%), VoxPopuli-Cleaned-AA (25%), and Earnings22-Cleaned-AA (25%). Speed Factor is the median over the last 7 days on 10-minute files. Price is AA's normalized USD per 1,000 minutes. Five entries are listed at $0. Each point is one model on one tested provider and endpoint. Methodology: https://artificialanalysis.ai/methodology/speech-to-text.
Hugging Face Open ASR Leaderboard multilingual raw results
Raw CSV revisions selected by the Space registry dated 2026-10-17. File revisions appear in measurement evidence. FLEURS, Common Voice (MCV), and MLS use test splits from hf-audio/open-asr-leaderboard-multilingual-datasets. Common Voice version is not specified by these CSVs. Hindi uses Voice Arena Monsoon conversational audio. These results use the leaderboard normalization and can differ from model cards. Models with a streaming-only label are omitted because the CSV does not state the streaming protocol. API version and date labels are preserved in evidence. Open-weight model metadata describes the original creator, not the Hugging Face account used to host a conversion.
Hugging Face Open ASR English raw results, 17 October 2026
Independent benchmark. Pinned snapshot selected by the official Space version registry. The stored avg column is preserved, not recomputed from the revised datasets or the rendered UI selection. Cleaned and original references remain separate datasets. TED-LIUM is absent in this snapshot. Short-form open-weight jobs use one NVIDIA H200 141 GB per the repository README after 24 July 2026. API model hardware is undisclosed. These are whole-file evaluations; no separate streaming track is present. A streaming-capable model name alone does not establish a live-audio test. Model aliases normalize creator names, underscores, Qwen -hf runtime suffixes, and the Voxtral 2602 API alias; original row IDs remain in evidence. Method: https://github.com/huggingface/open_asr_leaderboard/blob/main/README.md
Hugging Face Open ASR original eight-dataset English snapshot
Independent benchmark. Repository calls scripts/data the 26 March 2026 snapshot; this file last changed on 22 May 2026. Preserve the reported Avg. WER rather than recompute it. Original AMI, Earnings22, GigaSpeech, LibriSpeech clean/other, SPGISpeech, TED-LIUM, and VoxPopuli references differ from the 2026-10 cleaned track. -1 means unavailable and is omitted. Open-model hardware is NVIDIA A100-SXM4-80GB, driver 560.28.03, CUDA 12.6 per the contemporary README: https://github.com/huggingface/open_asr_leaderboard/blob/541b62c74abe3d14861fb90f8c77931dfff38a0a/README.md . All rows are whole-file evaluation, including streaming-capable architectures.
Coval independent STT seven-day aggregate API
Independent evaluation provider. Exact public API response read on 20 October 2026. Rolling seven-day window; results can change. Corpus WER uses pooled_value, not the mean of clip WER. Datasets are stt-v3 plus WildASR clean, accent, clipping, farfield, noisegap, and phonecodec. Runner is in us-east-1; provider regions differ. TTFS is final-transcript timestamp minus a shared VAD end-of-speech anchor, excluding connection setup and provider-specific endpointing. TTFT starts at the measured audio request rather than detected speech onset; mapped to latency_first_ms as the nearest metric and marked in each evidence string. Seconds convert to milliseconds. RTF includes paced live audio, so it is not imported as compute throughput. Missing metrics are omitted. API/weights identities are retained separately where exact checkpoint equivalence is not stated. Methods: https://github.com/coval-ai/benchmarks and https://benchmarks.coval.ai/benchmarks/time-to-final-segment .
Hugging Face Open ASR long-form raw results
Independent long-form track. Earnings21, Earnings22, TED-LIUM and CORAAL. Both Average and Avg (without CORAAL) are stored values, not recomputed. -1 is unavailable. The table has no per-row hardware column. Open-model runs follow the legacy A100-SXM4-80GB evaluation protocol; long-form has not migrated to the H200 Jobs protocol according to the current repository README. API hosts are providers, not disclosed GPUs. No streaming track is exposed.
Speechmatics official undiscounted API pricing
Official vendor pricing, embedded configuration rates.prerecorded.baseHourly and rates.realtime.baseHourly. The page initially enables training and commitment discounts; base rates here exclude both. modelTrainingMultiplier=0.67 and commitmentPreviewMultiplier=0.75 are not applied. Melia 1 maps to the shared speechmatics/melia model ID; Linden 1 maps to agent-stt-linden-1. Conversion: USD/hour × 1000/60.
Deepgram official API pricing
Official vendor list price, pay as you go. Current promotional streaming rates are recorded and identified in evidence; regular rates are not mixed with current prices. Monolingual and multilingual tariffs are separate observations of Nova-3, not separate trained-model IDs. Base transcription only, excluding add-ons and credits. Conversion: USD/minute × 1000.
AssemblyAI official API pricing
Official vendor list price. Base transcription only; credits, diarization, keyterms and other optional extras excluded. Streaming is billed by session duration. Conversion: USD/hour × 1000/60. Current Universal-3.6 Pro Realtime is distinct from older realtime Pro models and older Universal-3 Pro batch.
OpenAI official transcription pricing
Official vendor pricing. GPT Transcribe USD 0.0045/minute and live/realtime-whisper USD 0.017/minute are duration tariffs. GPT-4o and Mini are token-billed; the page labels USD 0.006/minute and USD 0.003/minute as estimated cost, so those minute equivalents are omitted under the no-estimates rule. Realtime speech-to-speech token prices are not treated as STT-only minute prices. Conversion: USD/minute × 1000.
ElevenLabs official API pricing
Official vendor published usage rates. Scribe v2 is USD 0.22/hour; Scribe v2 Realtime is USD 0.39/hour. Excludes keyterm prompting, transcript editing and taxes; plan-level terms can apply. Conversion: USD/hour × 1000/60.
Google Cloud Speech-to-Text V2 official pricing
Official vendor list price. V2 standard recognition, first 500000 minutes/month, USD 0.016/minute; channels billed separately and requests round up to one second. Batch and streaming standard-recognition rates are recorded. Dynamic-batch USD 0.003/minute is omitted because the pricing footnote does not explicitly establish Chirp 3 eligibility. Later marginal volume tiers are not mixed with first-tier rates.
AWS Transcribe official US East N. Virginia price file
Official vendor machine-readable list price, publication 2026-09-11, terms effective 2026-06-01. Standard TranscribeAudio USD 0.0001/second and StreamingAudio USD 0.0001667/second, 0-Inf. Excludes medical, call analytics and custom-model SKUs. The exact streaming SKU produces USD 10.002 per 1000 minutes, whereas the marketing page rounds to USD 0.01/minute. Conversion: USD/second × 60000. No model version is published.
Groq official speech-to-text API pricing
Official host list price for file transcription. The trained model creator remains OpenAI; host is Groq. Minimum billed audio duration is ten seconds per request. Conversion: USD/hour × 1000/60.
Meta Omnilingual ASR 7B LLM per-language CER resultsVendor report
Vendor self-report, not an independent evaluation. Original 7B LLM ASR model, not the later v2 model. Raw CSV contains 1,683 language-script rows and reports CER for all languages. The CSV does not identify the corpus, split, or aggregation for each row. Dataset is therefore labelled as the unspecified per-language evaluation suite, not as FLEURS. For languages with multiple scripts, the dataset label retains the script to prevent merging distinct numbers. Source language-script codes remain in evidence. ISO 639-3 base codes are canonicalized to two-letter BCP-47 codes when available; Mandarin cmn is shown as zh and Standard Arabic arb as ar. Do not compare this mixed evaluation suite directly with FLEURS, Common Voice, or WER results.
Qwen3-ASR Technical Report v2Vendor report
Vendor self-report, not an independent evaluation. Appendix A, Table A.2 provides per-language values. Section 4.1 uses CER for character-based languages and WER for word-delimited languages despite generic WER table headings. CER is retained for Chinese, Cantonese, Japanese, Korean, and Thai. Common Voice version is not stated. The zh_tw row is a distinct Taiwan subset with base language zh. MLS has a row labelled da although MLS normally has Dutch, not Danish; this ambiguous row is excluded. Appendix A.2 reports MLS-en Qwen3-ASR-0.6B as 6.04 while Table 3 reports 6.00; the appendix value is retained. Table 3 supplies competitor results run by Qwen on the test sets. These protocols differ from the HF leaderboard.
Argmax WhisperKit raw device performance resultsVendor report
Vendor self-report by the WhisperKit runtime author. Latest available run per selected exact model configuration and device, not latest OS pooled with old runs. Raw observations date from 2024-2025, not 2026; run timestamp, OS and runtime commit are in evidence. Includes standard Base EN, Small EN, Large v3, and Large v3 Turbo, plus the explicitly identified 632 MB compressed Turbo configuration on devices. The /turbo runtime suffix indicates an Argmax execution configuration and remains in evidence, not the creator/model ID. Test files are random 10-minute LibriSpeech and Earnings22 subsets; all are batch file runs, despite short-form being a proxy for streaming. End-to-end speed is audio seconds divided by wall seconds. Stored averages are preserved without recomputation. Method: https://huggingface.co/spaces/argmaxinc/whisperkit-benchmarks/blob/main/constants.py .
Microsoft MAI-Transcribe-1 model card, April 2, 2026Vendor report
Vendor self-report from Microsoft, not an independent evaluation. Page 2, Raw WER Data table embedded in the PDF. All 25 language columns are transcribed as printed. BR maps to Portuguese pt and NB to Norwegian Bokmal nb. The card calls every value WER, including Japanese, Chinese, and Thai, so this source retains WER and does not convert it to CER. The Whisper row is identified as Whisper-large-V3 in the page text. The Gemini label is Gemini 3.1 Flash as printed, not Flash-Lite. This older model-card comparison is separate from the later MAI-Transcribe-1.5 and MAI-Transcribe-2 cards, which give multilingual averages only.
Picovoice open speech-to-text benchmark, EnglishVendor report
Vendor self-report. Picovoice compares its own Cheetah/Leopard to competitors. English WER covers LibriSpeech test-clean/other, TED-LIUM and Common Voice. Word emission latency is mean delay from the end of each word to its emitted transcription, on 100 random LibriSpeech test-clean files. It is mapped to latency_first_ms as the nearest metric, not a first-partial-from-speech-start measurement. The README states AMD Ryzen 9 5900X, Ubuntu 22.04, 64 GB RAM, ten simultaneous cores for the separate core-hour experiment; accuracy and latency hardware are not separately specified, so local host is omitted. Core-hours are not imported as speed_factor. Engine source confirms English Whisper tiny/base/small/medium use .en weights, the Vosk model versions, and that Moonshine Tiny/Small/Medium rows use streaming architectures: https://github.com/Picovoice/speech-to-text-benchmark/blob/master/engine.py . API model versions are not disclosed.
NVIDIA Parakeet TDT 0.6B v3 model cardVendor report
Vendor self-report, not an independent evaluation. Structured model-index test results. WER excludes punctuation and capitalization errors. Parakeet uses greedy Transducer decoding without an external language model. CoVoST2 is a Common Voice-derived test corpus and is kept separate from Common Voice. Model-card values differ from the Hugging Face leaderboard normalization.
NVIDIA Canary 1B v2 model cardVendor report
Vendor self-report, not an independent evaluation. Structured model-index test results. WER excludes punctuation and capitalization errors. Parakeet uses greedy Transducer decoding without an external language model. CoVoST2 is a Common Voice-derived test corpus and is kept separate from Common Voice. Model-card values differ from the Hugging Face leaderboard normalization.
Mistral Voxtral Transcribe 2 release benchmarksVendor report
Vendor self-report from Mistral, not an independent evaluation. Exact numeric labels from the two per-language FLEURS charts. Nine languages are plotted even though the article discusses 13 supported languages. The charts say WER, including Chinese; WER is retained without conversion. Realtime uses the 480 ms transcription-delay setting recommended by the model card. The other streaming delay settings are omitted. Chinese competitor bars without numeric labels are omitted. Assembly Universal and Deepgram Nova have no model version in the chart and are excluded. Split is not specified.
Moonshine v2 paper, response latency and English WERVendor report
Vendor self-report by the model authors. Table 2 response latency is time from VAD-detected utterance end to returned final text on Apple MacBook M3, CPU execution. It excludes VAD decision delay and maps to latency_final_ms. Whisper uses faster-whisper and VAD-wrapped utterance processing. Original Moonshine Tiny/Base also use segmented realtime application processing; streaming mode here describes the evaluation, not a native streaming encoder. Table 3 WER is a separate offline Open ASR Transformers evaluation; it does not share the CPU latency run or hardware. Compute-load percentages and architecture lookahead are not imported as latency or speed. No latency statistic is explicitly specified by the paper, so stat is omitted.
FluidAudio Apple Silicon ASR benchmarksVendor report
Vendor self-report by the runtime/model-conversion author. Includes M4 Pro 48 GB macOS 26.0 and explicitly separate M5 Pro experiments. Mean per-file WER differs from corpus WER; preserve statistic labels. Neural Engine Core ML inference, not NVIDIA GPU runs. Per-language FLEURS rows are excluded from this file. Compilation, diarization and TTS results are excluded. Streaming Unified runs use chunked attention, not wall-clock-paced live audio, so RTFx is compute throughput. The v3 family section says macOS 27 while the Unified section says macOS 26; keep these separate in evidence. A 149-152x range is omitted rather than averaged. Phonon-2 is created by Fermion Research. Ultra is moondream/parakeet-ultra converted to FluidInference/parakeet-ultra-coreml, as established by Documentation/ASR/ParakeetUltra.md at the same repository revision.
Gladia speech-to-text benchmarks, MLS by languageVendor report
Vendor self-report from Gladia, not an independent evaluation. Audiobooks section, WER by Language HTML table. MLS covers German, Spanish, French, Italian, and Portuguese. Split and sampled subset are not specified. The rendered MLS by-language table labels the newer AssemblyAI model only as Universal v3, so that row is excluded because its exact version is unclear. Soniox v4 lacks async or realtime identification and is excluded. Deepgram v3 is identified as Nova-3 in the adjacent MLS summary. Only absolute per-language WER values are recorded; relative improvement percentages for Solaria-3 are not converted.
Argmax WhisperKit raw English quality resultsVendor report
Vendor self-report by the WhisperKit runtime author. JSONL despite the .json extension. Uncompressed standard model configurations only. LibriSpeech test set and a random 12-hour Earnings22 subset; full quality runs use Apple M2 Ultra Mac Studios. The stored average_wer is preserved. openai/whisper-large-v3-v20240930 is the 2024-09-30 Whisper Large v3 Turbo checkpoint, not a separate creator. Method and hardware: https://huggingface.co/spaces/argmaxinc/whisperkit-benchmarks/blob/main/constants.py .
Deepgram Nova-3 Multilingual code-switching updateVendor report
Vendor self-report, not an independent evaluation. Exact New Model numeric labels in batch and streaming code-switching charts. Previous Model values are excluded. Language labels denote code-switching datasets and must not be treated as monolingual FLEURS. Dataset names, sample counts, and splits are not disclosed. Some streaming language scores regress despite the article statement that all improve.
IBM Granite Speech 4.1 2B NAR model cardVendor report
Vendor self-report, not an independent evaluation. Additional results table. Greedy decoding, bfloat16 inference, jiwer after Whisper EnglishTextNormalizer. IBM explicitly warns that Open ASR Leaderboard results differ due to normalization and scoring. Common Voice version is 15. The model-card table does not state the split for these additional results. Other Granite 4.1 charts have unlabeled bars, so no values were estimated from them. Granite Speech 5.0 TurboCTC is English-only.
Argmax comparison of Apple SpeechAnalyzer and WhisperKitVendor report
Vendor self-report by Argmax, including its own free and paid SDK. Published 20 June 2025. Random 10% Earnings22 subset, about 12 hours. All tests use M4 Mac mini, macOS 26 Beta Seed 1. Speed factor is input audio seconds per wall second. Apple SpeechTranscriber is the model behind SpeechAnalyzer. The source labels Parakeet as nvidia/parakeet-v2; normalized to the 0.6B v2 checkpoint family. These are long-form offline runs, even though both frameworks also support realtime.
Pamir AI Distiller CM5 STT benchmarkVendor report
Hardware-vendor self-report by Pamir AI, updated 16 May 2025. Distiller CM5 is a Raspberry Pi CM5 module, four Cortex-A76 cores at 2.4 GHz, 8 GB LPDDR4, CPU only. 250 Common Voice Delta 20.0 validated clips. Speed factor is the exact reciprocal of the reported rounded RTF, not an independent observation. Only rows with a model size/version stated are included; unspecified Vanilla/Fast/OpenVINO Whisper rows and the ambiguous 0.11B Parakeet checkpoint are omitted. Quantization/runtime remain in evidence rather than model IDs.
parakeet-mlx maintainer benchmark on MacBook Air M2Vendor report
Runtime-author self-report, 16 May 2025. parakeet-mlx 0.2.7, mlx-community/parakeet-tdt-0.6b-v2, bf16, 120-second chunks and 15-second overlap. MacBook Air M2 16 GB 2022, fanless and thermally throttled. Three runs per file. No WER is supplied. The discussion includes later commenter hardware; only the maintainer original table is imported.
Kyutai STT release pageVendor report
Vendor self-report. The 500 ms and 2.5 s delays are configured text-stream alignment delays, not measured end-of-speech or first-partial latencies. They are mapped to latency_first_ms as the nearest metric and marked in evidence. The release also discusses faster-than-realtime flushing, which is not included as a measured latency. GPU stream concurrency is not a single-stream speed factor. English WER and throughput for STT 2.6B are recorded from the independent HF snapshot.