# Speech recognition model manifest. # # Ordered list of models that can perform speech recognition. The first model # matching a BCP-47 tag is selected. # # Fields # ------ # id : short identifier used in generated symbol names and in the pref # media.webspeech.recognition.model. # supported_locales : Locale tags this artifact recognizes, spelled as the model # expects them; SpeechModelFor() negotiates a request against # them, returning the model id the utility process passes around # plus that locale (SpeechRecognitionModelMapping.h). # streaming : true = cache-aware streaming RNN-T/transducer (parakeet_capi_stream_*) # false = offline CTC batch path (parakeet_capi_transcribe_pcm) # repo : ModelHub repository (org/name) # filename : GGUF file name inside the repository # revision : ModelHub revision tag # size_mb : Approximate download size shown in the permission prompt (integer MB) # quant : Quantization identifier (informational) # latency_ms : Approximate per-chunk inference latency in ms; 0 = not yet measured # display_name : Human-readable model name shown in about:addons' "Manage # On-Device AI Models" panel. Several distinct upstream models are # rehosted under the single "speech-recognition/parakeet" ModelHub # bucket (repo), so that bucket's path cannot be used to derive a # precise name; this is the real per-artifact name instead. # hf_url : Upstream Hugging Face model card for this artifact, shown as the # "View on Hugging Face" link. Not derivable from repo/revision # (see display_name) since the ModelHub bucket path and the real # upstream Hugging Face repo path differ for this feature. models: # --- English streaming model (EOU-aware, RNN-T transducer) --- - id: english supported_locales: - en streaming: true repo: "speech-recognition/parakeet" filename: "realtime_eou_120m-v1-q5_k.gguf" revision: "main" size_mb: 141 quant: "q5_k" latency_ms: 0 display_name: "Parakeet Realtime EOU 120M v1" hf_url: "https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1" # --- English TDT-CTC 110m (offline CTC; multiple quants) --- - id: english_tdt_ctc_q4_k supported_locales: - en streaming: false repo: "speech-recognition/parakeet" filename: "tdt_ctc-110m-q4_k.gguf" revision: "main" size_mb: 131 quant: "q4_k" latency_ms: 0 display_name: "Parakeet TDT-CTC 110M" hf_url: "https://huggingface.co/nvidia/parakeet-tdt_ctc-110m" - id: english_tdt_ctc_q5_k supported_locales: - en streaming: false repo: "speech-recognition/parakeet" filename: "tdt_ctc-110m-q5_k.gguf" revision: "main" size_mb: 143 quant: "q5_k" latency_ms: 0 display_name: "Parakeet TDT-CTC 110M" hf_url: "https://huggingface.co/nvidia/parakeet-tdt_ctc-110m" - id: english_tdt_ctc_q6_k supported_locales: - en streaming: false repo: "speech-recognition/parakeet" filename: "tdt_ctc-110m-q6_k.gguf" revision: "main" size_mb: 156 quant: "q6_k" latency_ms: 0 display_name: "Parakeet TDT-CTC 110M" hf_url: "https://huggingface.co/nvidia/parakeet-tdt_ctc-110m" - id: english_tdt_ctc_q8_0 supported_locales: - en streaming: false repo: "speech-recognition/parakeet" filename: "tdt_ctc-110m-q8_0.gguf" revision: "main" size_mb: 178 quant: "q8_0" latency_ms: 0 display_name: "Parakeet TDT-CTC 110M" hf_url: "https://huggingface.co/nvidia/parakeet-tdt_ctc-110m" - id: english_tdt_ctc_f16 supported_locales: - en streaming: false repo: "speech-recognition/parakeet" filename: "tdt_ctc-110m-f16.gguf" revision: "main" size_mb: 268 quant: "f16" latency_ms: 0 display_name: "Parakeet TDT-CTC 110M" hf_url: "https://huggingface.co/nvidia/parakeet-tdt_ctc-110m" # --- Multilingual (nemotron, streaming RNN-T) --- # Last, so English goes to the English models above even though this one # also lists en-US/en-GB. - id: multilingual # Straight from the model's own prompt dictionary # (parakeet.prompt.dictionary.keys), minus its malformed keys (enGB, esES) and # its "auto" default, which is what passing no locale at all selects. # The model card claims quality for a subset, in decreasing order: # core: en-US en-GB es-US es-ES fr-FR fr-CA it-IT pt-BR # pt-PT nl-NL de-DE tr-TR ru-RU ar-AR hi-IN ja-JP # ko-KR vi-VN uk-UA # broad coverage: pl-PL sv-SE cs-CZ nb-NO da-DK bg-BG fi-FI hr-HR # sk-SK zh-CN hu-HU ro-RO et-EE # adaptation-ready: el-GR lt-LT lv-LV mt-MT sl-SI he-IL th-TH nn-NO # the rest the model accepts but the card does not claim. supported_locales: - en-US - en - en-GB - es-ES - es-US - es - zh-CN - zh-ZH - zh-TW - hi-IN - hi - hi-HI - ar-AR - ar - fr-FR - fr - de-DE - de - ja-JP - ja-JA - ru-RU - ru - pt-BR - pt-PT - pt - ko-KR - ko - ko-KO - it-IT - it - nl-NL - nl - pl-PL - pl - tr-TR - tr - uk-UA - uk - ro-RO - ro - el-GR - el - cs-CZ - cs - hu-HU - hu - sv-SE - sv - da-DK - da - fi-FI - fi - no-NO # Quoted: YAML reads a bare no as false. - "no" - nb-NO - nb - nn-NO - nn - sk-SK - sk - hr-HR - hr - bg-BG - bg - lt-LT - lt - et-EE - et - lv-LV - lv - sl-SI - sl - th-TH - vi-VN - id-ID - ms-MY - bn-IN - ur-PK - fa-IR - ta-IN - te-IN - mr-IN - gu-IN - kn-IN - ml-IN - si-LK - ne-NP - km-KH - sw-KE - am-ET - ha-NG - zu-ZA - yo-NG - ig-NG - af-ZA - rw-RW - so-SO - ny-MW - ln-CD - or-KE - he-IL - ku-TR - az-AZ - ka-GE - hy-AM - uz-UZ - tg-TJ - ky-KG - qu-PE - ay-BO - gn-PY - nah-MX - mi-NZ - haw-US - sm-WS - to-TO - fr-CA - mt-MT streaming: true repo: "speech-recognition/parakeet" filename: "nemotron-3.5-asr-streaming-0.6b-q5_k.gguf" revision: "main" size_mb: 785 quant: "q5_k" latency_ms: 0 display_name: "Nemotron 3.5 ASR Streaming 0.6B" hf_url: "https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b"