89 audio models from 30 providers, sorted into six tasks. Most popular for speech to text: Granite Speech 5.0 470M TurboCTC NC (IBM). For speech output: Kokoro-82M (hexgrad).
providers
30
with open weights
22
with a published price
38
Which audio model for which task?
The three most popular models per task. Each task name leads to its section below.
Speech to text32 modelsAudio becomes text: transcription, subtitles, dictation.
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
GraduallyBench Audio
Two tests of our own: how many words does a speech recognizer get wrong, and how completely does a text come through in speech output?
Speech recognition
20models with a German and English value, 8 more in the table
Speech output
20models with a German and English value, 3 more in the table
Technical tests only
24music, dialogue, translation and more
Data as of
11/10/2026evaluated automatically
Speech recognition
Up to 36 read recordings in German and English; not every model covers both languages. The word error rate (WER) counts wrong, missing and extra words per 100 words of the reference, the character error rate (CER) does the same for characters.
Speech recognition word error rate
Misrecognized words in percent (German and English), lower is better
Up to 32 texts in German and English. The generated speech is transcribed back to text with GPT Transcribe and compared with the source. This measures whether the text comes through completely, not how natural the voice sounds.
Speech output text comparison
Differing words after re-transcription in percent (German and English), lower is better
3 more models have no complete German and English value and are not in the chart. Smaller test scope: Fish Audio S2 Pro, Fish Audio S2.1 Pro, MAI-Voice-2, MAI-Voice-2-Flash, Sonic 3.6, Voxtral Mini TTS, GPT-4o Mini TTS.
As of 11/10/2026. Source: GraduallyBench Audio
Sorted by the German and English word error rate; models without this value come last.
What is availableDialogue tasks after exclusions: 50.0 % (3 / 6 assessable dialogue cases). Only a technical test is available. The selection of test inputs is limited.
What is availableDialogue tasks after exclusions: 83.3 % (5 / 6 assessable dialogue cases). Only a technical test is available. The selection of test inputs is limited.
What is availableDialogue tasks after exclusions: 100.0 % (5 / 5 assessable dialogue cases). The test coverage is incomplete. Only a technical test is available.
What is availableWord distance from the translation reference: 72.5 % (37 / 51 reference words). Only a technical test is available. The selection of test inputs is limited.
What is availableDiarization, technical pretest: 15.8 % (67.32 / 426.42 seconds reference speaker time). Only a technical test is available. The selection of test inputs is limited.
The functional test does not assess content quality.
Only a technical test is available.
The selection of test inputs is limited.
The route does not confirm the exact model version.
Show test details
Technical measurement
Public demo · Language not documented
1 valid output from 1 attempted and 1 planned test slots.
Only a short functional test is available.
No further scored metric is available for this test.
Technical measurement
Public demo · Chinese, English, Spanish
2 valid outputs from 3 attempted and 3 planned test slots.
The target language of the output was not confirmed.
Chinese identifies the submitted input. English and Spanish are requested target languages. The input audio identity and the actual language of the output audio were not independently confirmed.
No further scored metric is available for this test.
The functional test does not assess content quality.
Only a technical test is available.
The selection of test inputs is limited.
The route does not confirm the exact model version.
Show test details
Technical measurement
Public demo · Language not documented
1 valid output from 1 attempted and 1 planned test slots.
Only a short functional test is available.
No further scored metric is available for this test.
Technical measurement
Public demo · Chinese, English, Spanish
2 valid outputs from 3 attempted and 3 planned test slots.
The target language of the output was not confirmed.
Chinese identifies the submitted input. English and Spanish are requested target languages. The input audio identity and the actual language of the output audio were not independently confirmed.
No further scored metric is available for this test.
The tested weights have noncommercial usage terms.
The functional test does not assess content quality.
Show test details
Technical measurement
Local · Language not documented
Execution Apple Metal. Precision Q4_0 · Decoder precision FP16.
2 valid outputs from 2 attempted and 2 planned test slots.
No further scored metric is available for this test.
WER compares words and CER compares characters with the reference. The speech output was re-transcribed with GPT Transcribe. Its exact backend revision is unconfirmed. Read speech and digital silence do not cover spontaneous speech or real background noise. Error rates can exceed 100 %. Where used, the test references were not independently reviewed by humans. Test coverage and methods differ by model. Each metric stands on its own; there is no combined overall score across tests. 12 additional models across Decision and Audio have no usable output and are omitted. Data compiled on 11/10/2026. How we test · Download all Decision and Audio results as JSON
First published cohorts
GraduallyBench STT v1
Speech recognition compared automatically.
The pilot uses the same 32 read-speech recordings in German and English, plus four digital-silence clips. The comparison uses source transcripts without human verification. Each language contains four speakers. Spontaneous speech and background noise are outside this cohort.
Word error rate (WER) counts replaced, omitted and inserted words relative to the reference words. Character error rate (CER) compares characters. Lower values mean fewer differences from the reference. WER can exceed 100 %. Counts below each rate show errors / reference units. Silence is evaluated separately.
Protocol audio-stt-read-speech-automated-v1. Actual provider billing is unknown.
GPT Transcribe
OpenAI · gpt-transcribe · https://api.openai.com/v1/audio/transcriptions · Tested Nov 5, 2026
Successful requests 36 / 36. The exact backend version is not confirmed by the requested route.
Read-speech transcription
Language
Successful responses
WER
CER
WER, speaker mean
German
16 / 16
6.73 %37 / 550
2.42 %79 / 3262
6.59 %n = 4 / 4
English
16 / 16
1.63 %8 / 490
0.81 %21 / 2580
1.77 %n = 4 / 4
Digital silence
Language setting
Clips with invented speech
Invented words
DE
0 / 2
0
EN
0 / 2
0
Request duration, median / p95 1,154 / 1,974 ms (n = 36). Includes network time and client processing.
Deepgram Nova 3
Cloudflare Workers AI · @cf/deepgram/nova-3 · Tested Nov 5, 2026
Successful requests 36 / 36. The exact backend version is not confirmed by the requested route.
Read-speech transcription
Language
Successful responses
WER
CER
WER, speaker mean
German
16 / 16
8.55 %47 / 550
2.61 %85 / 3262
8.16 %n = 4 / 4
English
16 / 16
2.04 %10 / 490
0.66 %17 / 2580
2.02 %n = 4 / 4
Digital silence
Language setting
Clips with invented speech
Invented words
DE
0 / 2
0
EN
0 / 2
0
Request duration, median / p95 1,180.5 / 2,849 ms (n = 36). Includes network time and client processing.
Scribe v2
fal.ai · fal-ai/elevenlabs/speech-to-text/scribe-v2 · Tested Nov 5, 2026
Successful requests 36 / 36. The exact backend version is not confirmed by the requested route.
Read-speech transcription
Language
Successful responses
WER
CER
WER, speaker mean
German
16 / 16
3.64 %20 / 550
0.98 %32 / 3262
3.27 %n = 4 / 4
English
16 / 16
1.63 %8 / 490
0.62 %16 / 2580
1.68 %n = 4 / 4
Digital silence
Language setting
Clips with invented speech
Invented words
DE
0 / 2
0
EN
0 / 2
0
Request duration, median / p95 2,745 / 6,215 ms (n = 36). Includes network time and client processing.
GraduallyBench TTS v1
Automatic text comparison after speech generation.
Each first audio output is fully decoded and transcribed once by gpt-transcribe. Word error rate (WER) and character error rate (CER) compare this transcript with the frozen input script. Lower values mean fewer text differences. Errors can also come from the transcription model.
This model-dependent text proxy does not measure naturalness, voice preference or human intelligibility. Numbers and names depend on the fixed normalization. The requested voice IDs are recorded; the provider does not independently confirm the delivered voice identity.
Protocol audio-tts-asr-content-proxy-v1. Actual provider billing is unknown.
Analysis date Nov 5, 2026
Eleven v4
fal.ai · elevenlabs/tts/eleven-v4
Generated
32
Fully decoded
32 / 32
Transcribed once
32 / 32
One already generated original was secured through a separate download.
No single model covers every task, so the hub ranks each task separately. By popularity the leaders are currently: Speech to text: Granite Speech 5.0 470M TurboCTC NC; Speech output: Kokoro-82M; Real-time voice dialog: Amazon Nova 2 Sonic; Speech translation: Index-Echo-S2ST-9B; Music and sound: MiniMax-Music3; Audio analysis: Nemotron-3-Diarization. Popularity measures attention and use, not audio quality. A shared quality benchmark across all models of a task is not available yet, so test two or three candidates with your own recordings.
4 speech output models publish open weights, among them Kokoro-82M, Fish Audio S2 Pro and Maya 2 Native. You can run open models yourself without API fees, but you need suitable hardware, and the license decides commercial use. Hosted voices usually charge per character, token or minute.
12 models in the hub take speech in and answer with speech, which voice agents and phone bots need. The most popular are Amazon Nova 2 Sonic, Eleven Realtime v1 Max and Eleven Realtime v1 Mini. Response time and turn-taking decide how natural a conversation feels; compare them in your own setup before you build on one model.
22 of 89 audio models list local deployment on their profile, for example Kokoro-82M, Nemotron-3-Diarization and MiniMax-Music3. Use the "Runs locally" filter in the table for the full list. Check model size and license on each profile.
38 of 89 models have a published price on their profile. Units differ: per minute or hour of audio, per million characters, per million tokens or in credits, so the table shows each price as the provider states it. Examples: Voxtral Mini TTS: $16 per 1M characters; Amazon Nova 2 Sonic: Separate token rates for audio and text input and audio and text output; text rates also apply to transcription, tool calls, grounding, and history. Standard, Priority, and Flex tiers must be checked separately..
The hub sorts every model into one of six tasks: Speech to text (32), Speech output (23), Real-time voice dialog (12), Speech translation (8), Music and sound (11), Audio analysis (3). Each model has exactly one task, chosen from its documented input and output.
Changelog
The latest updates and improvements to our AI audio model comparison
November 4, 2026v1.0
Initial release
Published 89 audio models, each assigned to one of six tasks
Added popularity charts per task, a searchable table with filters and an FAQ