Parakeet TDT 0.6B v3 is NVIDIA’s 600-million-parameter multilingual speech-recognition model. It expands the English-only v2 model to 25 European languages and automatically detects the language of the audio without a separate language prompt. The result is a transcript of the speech, with punctuation and capitalization.
The supported languages are Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian and Ukrainian. This list belongs to the v3 edition; other Parakeet decoder types and language variants are separate models.
The architecture combines a FastConformer encoder with a token-and-duration transducer decoder. The encoder turns the acoustic signal into representations; the decoder predicts text tokens together with the durations needed to advance through the audio frames. This TDT behavior distinguishes it from the family’s CTC and ordinary RNN-T decoders.
Prepare single-channel audio sampled at 16 kHz. The maker’s direct input formats are WAV and FLAC, and its implementation examples use a processor to prepare the waveform for recognition. Keep the recording identifier and original audio alongside the returned text so that a transcript can be checked or aligned with the source.
Timestamping adds timing information to the recognized speech. NVIDIA’s NeMo example returns character, word and segment timings when timestamp generation is enabled. A subtitle or archive application can use segment boundaries to connect a passage with the recording, or word timings to locate a particular phrase. The transcript and its timing fields are different parts of the result and should remain associated.
For long recordings, the attention configuration matters. The maker describes up to 24 minutes with full attention on an 80 GB A100, and up to three hours with local attention. Its long-form example switches the encoder to local attention with bounded left and right context. Those durations describe specific producer configurations, rather than one unconditional limit for every deployment.
The manufacturer also publishes a chunked-inference workflow for streaming use. Its example defines chunk length and left and right audio context around each chunk. This is a distinct processing recipe from the native cache-aware Nemotron streaming model. An application should retain the chosen recipe and the original timing reference when collecting a recording’s transcript across sections.
The training mix includes 10,000 hours of human-transcribed speech and 660,000 hours of pseudo-labeled Granary data, with punctuation and capitalization preserved. The producer also supports fine-tuning this released checkpoint on another speech dataset through its documented model workflow; the tuned checkpoint remains distinct from the base v3 model. The model’s text output is therefore suitable as the speech-to-text stage of a larger workflow. A meeting archive can attach the transcript to the recording; a search application can index the recognized passage; a captioning tool can use the returned timing metadata.
Recognition quality varies with language, accent, noise and the type of speech. The producer specifically notes that vocabulary not represented during training can be missed and that isolated words or incomplete sentences provide weaker context. Preserve the original audio for reviewing names and specialized terms rather than turning an uncertain recognition into a verified fact.
The useful artifact is the transcript with the requested timing information. Summaries, translations, speaker analysis and actions are separate application tasks, not features implied by a speech-to-text name. Keep the v3 identity and producer attention mode distinct from other checkpoints and from converted or quantized artifacts.
", "aboutText": "Parakeet TDT 0.6B v3 is NVIDIA’s 600-million-parameter multilingual speech-recognition model. It expands the English-only v2 model to 25 European languages and automatically detects the language of the audio without a separate language prompt. The result is a transcript of the speech, with punctuation and capitalization.\n\nThe supported languages are Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian and Ukrainian. This list belongs to the v3 edition; other Parakeet decoder types and language variants are separate models.\n\nThe architecture combines a FastConformer encoder with a token-and-duration transducer decoder. The encoder turns the acoustic signal into representations; the decoder predicts text tokens together with the durations needed to advance through the audio frames. This TDT behavior distinguishes it from the family’s CTC and ordinary RNN-T decoders.\n\nPrepare single-channel audio sampled at 16 kHz. The maker’s direct input formats are WAV and FLAC, and its implementation examples use a processor to prepare the waveform for recognition. Keep the recording identifier and original audio alongside the returned text so that a transcript can be checked or aligned with the source.\n\nTimestamping adds timing information to the recognized speech. NVIDIA’s NeMo example returns character, word and segment timings when timestamp generation is enabled. A subtitle or archive application can use segment boundaries to connect a passage with the recording, or word timings to locate a particular phrase. The transcript and its timing fields are different parts of the result and should remain associated.\n\nFor long recordings, the attention configuration matters. The maker describes up to 24 minutes with full attention on an 80 GB A100, and up to three hours with local attention. Its long-form example switches the encoder to local attention with bounded left and right context. Those durations describe specific producer configurations, rather than one unconditional limit for every deployment.\n\nThe manufacturer also publishes a chunked-inference workflow for streaming use. Its example defines chunk length and left and right audio context around each chunk. This is a distinct processing recipe from the native cache-aware Nemotron streaming model. An application should retain the chosen recipe and the original timing reference when collecting a recording’s transcript across sections.\n\nThe training mix includes 10,000 hours of human-transcribed speech and 660,000 hours of pseudo-labeled Granary data, with punctuation and capitalization preserved. The producer also supports fine-tuning this released checkpoint on another speech dataset through its documented model workflow; the tuned checkpoint remains distinct from the base v3 model. The model’s text output is therefore suitable as the speech-to-text stage of a larger workflow. A meeting archive can attach the transcript to the recording; a search application can index the recognized passage; a captioning tool can use the returned timing metadata.\n\nRecognition quality varies with language, accent, noise and the type of speech. The producer specifically notes that vocabulary not represented during training can be missed and that isolated words or incomplete sentences provide weaker context. Preserve the original audio for reviewing names and specialized terms rather than turning an uncertain recognition into a verified fact.\n\nThe useful artifact is the transcript with the requested timing information. Summaries, translations, speaker analysis and actions are separate application tasks, not features implied by a speech-to-text name. Keep the v3 identity and producer attention mode distinct from other checkpoints and from converted or quantized artifacts.", "editorialOverlay": { "copy_fields": null, "copy_sha256": null, "copy_source": null, "html_sha256": "1a68f3837f45ab81b82e256f6f338c9829339aaa8c4c982e498802f65c90801e", "module_sha256": "ee9dc1c50b1e6a8fb7fa062030fcb4646b9aaaeb08be513b723a5a92a7290f06", "raw_model_unchanged": true, "schema": "ailoot-native-model-editorial-v1" }, "locale": "en", "methods": [ { "html": "nvidia/parakeet-tdt-0.6b-v3NVIDIA’s Parakeet TDT 0.6B v3 is a 600M speech-to-text model for 25 European languages, with automatic language detection. It produces punctuated transcripts and supports word and segment timestamps, with long-audio modes tied to the producer’s attention configuration.
This description covers the model’s capabilities. Available in Ailoot inputs, outputs and options are listed for the selected method below.
Parakeet TDT 0.6B v3 is NVIDIA’s 600-million-parameter multilingual speech-recognition model. It expands the English-only v2 model to 25 European languages and automatically detects the language of the audio without a separate language prompt. The result is a transcript of the speech, with punctuation and capitalization.
The supported languages are Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian and Ukrainian. This list belongs to the v3 edition; other Parakeet decoder types and language variants are separate models.
The architecture combines a FastConformer encoder with a token-and-duration transducer decoder. The encoder turns the acoustic signal into representations; the decoder predicts text tokens together with the durations needed to advance through the audio frames. This TDT behavior distinguishes it from the family’s CTC and ordinary RNN-T decoders.
Prepare single-channel audio sampled at 16 kHz. The maker’s direct input formats are WAV and FLAC, and its implementation examples use a processor to prepare the waveform for recognition. Keep the recording identifier and original audio alongside the returned text so that a transcript can be checked or aligned with the source.
Timestamping adds timing information to the recognized speech. NVIDIA’s NeMo example returns character, word and segment timings when timestamp generation is enabled. A subtitle or archive application can use segment boundaries to connect a passage with the recording, or word timings to locate a particular phrase. The transcript and its timing fields are different parts of the result and should remain associated.
For long recordings, the attention configuration matters. The maker describes up to 24 minutes with full attention on an 80 GB A100, and up to three hours with local attention. Its long-form example switches the encoder to local attention with bounded left and right context. Those durations describe specific producer configurations, rather than one unconditional limit for every deployment.
The manufacturer also publishes a chunked-inference workflow for streaming use. Its example defines chunk length and left and right audio context around each chunk. This is a distinct processing recipe from the native cache-aware Nemotron streaming model. An application should retain the chosen recipe and the original timing reference when collecting a recording’s transcript across sections.
The training mix includes 10,000 hours of human-transcribed speech and 660,000 hours of pseudo-labeled Granary data, with punctuation and capitalization preserved. The producer also supports fine-tuning this released checkpoint on another speech dataset through its documented model workflow; the tuned checkpoint remains distinct from the base v3 model. The model’s text output is therefore suitable as the speech-to-text stage of a larger workflow. A meeting archive can attach the transcript to the recording; a search application can index the recognized passage; a captioning tool can use the returned timing metadata.
Recognition quality varies with language, accent, noise and the type of speech. The producer specifically notes that vocabulary not represented during training can be missed and that isolated words or incomplete sentences provide weaker context. Preserve the original audio for reviewing names and specialized terms rather than turning an uncertain recognition into a verified fact.
The useful artifact is the transcript with the requested timing information. Summaries, translations, speaker analysis and actions are separate application tasks, not features implied by a speech-to-text name. Keep the v3 identity and producer attention mode distinct from other checkpoints and from converted or quantized artifacts.
POST /v1/audio/transcriptionsaudiotext| Field | Type | Requirement | Values & limits | Default |
|---|---|---|---|---|
Fieldmodel | Typestring | RequirementRequired | Values & limits nvidia/parakeet-tdt-0.6b-v3 | Default— |
Fieldinput_audio | Typeobject | RequirementRequired | Values & limits | Default— |
Fieldlanguage | Typestring | RequirementOptional | Values & limits | Default— |
Fieldtemperature | Typenumber | RequirementOptional | Values & limits Minimum: 0 Maximum: 1 | Default— |
Fieldresponse_format | Typestring | RequirementOptional | Values & limits json, verbose_json | Defaultjson |
Fieldtimestamp_granularities | Typearray | RequirementOptional | Values & limits | Default— |
Fieldprovider | Typeobject | RequirementOptional | Values & limits | Default— |
| Parameter | Contract |
|---|---|
Parametercontent_type | Contractapplication/json |
Parameterrequired | Contractmodel, input_audio |
Parametermaterial_types | Contractaudio |
Parametermaterial_conditions | ContractRead an actual WAV file and encode its real bytes for the declared JSON input_audio; no multipart route is advertised for this exact JSON-only declaration. |
Lifecycle: synchronous_result_then_owned_library
GET /v1/media/tasks/{task_id} · actual returned task_id or media library record
GET /v1/media/assets/{asset_id}/content
GET /v1/media/assets
GET /v1/media/assets/{asset_id}
The displayed rates estimate the cost before generation. The final charge is calculated from the cost recorded for the original generation. Measured output duration and forecast quantities are not the final billable quantity.
Tariff references do not add request options. Supported parameters are listed for the selected method.
Match each charge to the conditions shown. Use one rate for each set of alternative configurations.
| Configuration | Rate schedule | Forecast quantity | Billing unit | Charge |
|---|---|---|---|---|
| ConfigurationBase token rates | Rate scheduleBase rate | Forecast quantity— | Billing unit— | Charge Input · Input tokens · Estimate for comparison only — / 1K tokens Output · Output tokens · Estimate for comparison only — / 1K tokens Input · Cache read · Estimate for comparison only — / 1K tokens Input · Cache write · Estimate for comparison only — / 1K tokens Input · Cache write, 1 hour · Estimate for comparison only — / 1K tokens |
| Configurationinput audio second | Rate schedule— | Forecast quantity— | Billing unitper input audio second | Charge Input · input audio second · Estimate for comparison only $0.00005 per input audio second |
| ConfigurationEstimated generation cost | Rate schedule— | Forecast quantityone generation estimate | Billing unit— | Charge Output · Estimated generation cost · Estimate for comparison only $0.00005 estimate before generation This estimate is not an additional charge or a fixed rate for the completed generation. |
Ailoot reserves an estimated amount before a request. After the result, it settles the final charge from measured usage. Streaming settlement uses the final usage record. How reservation and settlement work →
set -eu\n: "${AILOOT_API_KEY:?Set your Ailoot API key}"\n: "${AILOOT_AUDIO_FILE:?Set the path to your actual WAV recording}"\npython3 - <<'PY_REQUEST'\nimport json,os\nfrom pathlib import Path\nimport base64\naudio=Path(os.environ['AILOOT_AUDIO_FILE']).read_bytes()\nif not (audio[:4]==b'RIFF' and audio[8:12]==b'WAVE'): raise SystemExit('Supply an actual WAV audio file')\nif len(audio)>26214400: raise SystemExit('Use a documented supported large-file flow')\nrequest={'model': 'nvidia/parakeet-tdt-0.6b-v3', 'input_audio': {'data': '${BASE64_ACTUAL_WAV_BYTES}', 'format': 'wav'}}\nrequest['input_audio']['data']=base64.b64encode(audio).decode('ascii')\nPath('request.json').write_text(json.dumps(request,ensure_ascii=False),encoding='utf-8')\nPY_REQUEST\ncurl --fail-with-body --silent --show-error https://api.ailoot.ai/v1/audio/transcriptions \\\n -H "Authorization: Bearer $AILOOT_API_KEY" -H 'Content-Type: application/json' \\\n --data-binary @request.json -D response-headers.txt -o response.bin\n# Keep the complete original response and headers. Read every returned segment, word and usage field.\n# Preserve actual operation/task/asset IDs; continue reads of the original result without resubmitting.\nGET /v1/media/tasks/{task_id}GET /v1/media/assets/{asset_id}/contentGET /v1/media/assetsGET /v1/media/assets/{asset_id}set -eu\n: "${AILOOT_API_KEY:?Set your Ailoot API key}"\ncurl --fail-with-body --silent --show-error https://api.ailoot.ai/v1/media/assets \\\n -H "Authorization: Bearer $AILOOT_API_KEY" -o owned-assets.json\n# Preserve every returned asset and follow its exact signed content URL. Do not construct a signed token.\nKeep every complete returned response, file and content part.
Retain the complete response and headers, every actual output, original returned identifiers and all usage fields. Recover the original operation without submitting creation again.
Keep the returned X-Client-Request-ID for reconciliation. Read only actual returned task_id/asset_id values or the owned media library; never build a task URL from the operation ID and never resubmit creation.
| Field | Type |
|---|---|
Fieldtext | Type |
Fieldusage | Type |
Fieldsegments[] | Type |
Fieldwords[] | Type |
Fieldspeaker fields | Type |
Fieldtimestamps | Type |
Ailoot Ultra nvidia/parakeet-tdt-0.6b-v3
Use the method paths shown above.