Repository navigation
Export https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 to sherpa-onnx - #2500
Conversation
WalkthroughIntroduces v3 export and testing for NeMo Parakeet TDT 0.6b alongside v2, updates CI to run both versions with per-version artifact handling, disables HF transfer, adds macOS runner cleanup, adjusts v2 export to remove FP16 and use external encoder weights, adds APK/kotlin entries for v3, and makes minor C++ message/comment tweaks. Changes
Sequence Diagram(s)sequenceDiagram
actor Dev as GitHub Actions
participant Job as CI Job (Matrix v2/v3)
participant Script as run.sh (v2/v3)
participant Export as export_onnx.py
participant HF as HuggingFace Repo
Dev->>Job: Start matrix (version=v2,v3)
Job->>Job: macOS cleanup, set env
Job->>Script: Execute per-version run.sh
Script->>Export: Export encoder/decoder/joiner ONNX
Export-->>Script: ONNX + encoder.weights + tokens.txt
Script->>Script: Run ONNX tests (FP32 / INT8)
Job->>HF: Track LFS (*.onnx, *.weights) and push artifacts
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Possibly related PRs
Poem
Tip 🔌 Remote MCP (Model Context Protocol) integration is now available!Pro plan users can now connect to remote MCP servers from the Integrations page. Connect with popular remote MCPs such as Notion and Linear to add more context to your reviews and chats. ✨ Finishing Touches
🧪 Generate unit tests
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. CodeRabbit Commands (Invoked using PR/Issue comments)Type Other keywords and placeholders
CodeRabbit Configuration File (
|
There was a problem hiding this comment.
Pull Request Overview
This PR adds support for the newer version (v3) of NVIDIA's parakeet-tdt-0.6b model by exporting it to sherpa-onnx format. The v3 model supports 25 languages, maintaining the same usage pattern as v2 but with improved language coverage.
Key changes:
- Added export scripts and configuration for parakeet-tdt-0.6b-v3 model
- Updated Kotlin API to include the new model configuration (case 40)
- Fixed minor grammar issues in existing code comments
Reviewed Changes
Copilot reviewed 10 out of 10 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
| sherpa-onnx/kotlin-api/OfflineRecognizer.kt | Added model configuration for parakeet-tdt-0.6b-v3 (case 40) |
| sherpa-onnx/csrc/offline-tts.h | Fixed grammar in comment ("stop" -> "stops") |
| sherpa-onnx/csrc/offline-tts-model-config.cc | Fixed grammar in error message ("at exactly" -> "exactly") |
| scripts/nemo/parakeet-tdt-0.6b-v3/ | Added complete export pipeline for v3 model including test script, run script, and ONNX export |
| scripts/nemo/parakeet-tdt-0.6b-v2/ | Cleaned up v2 export scripts by removing fp16 support and unused dependencies |
| scripts/apk/generate-vad-asr-apk-script.py | Added v3 model to APK generation script |
| .github/workflows/export-nemo-parakeet-tdt-0.6b-v2.yaml | Updated CI workflow to support both v2 and v3 model exports |
Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.
| for w in en.wav de.wav fr.wav es.wav; do | ||
| echo "---fp32----" | ||
| python3 ./test_onnx.py \ | ||
| --encoder ./encoder.int8.onnx \ |
There was a problem hiding this comment.
The encoder path is inconsistent between fp32 and int8 tests. The fp32 test uses encoder.int8.onnx instead of encoder.onnx, which means it's not actually testing the fp32 model.
| --encoder ./encoder.int8.onnx \ | |
| --encoder ./encoder.onnx \ |
| @@ -0,0 +1 @@ | |||
| ../parakeet-tdt-0.6b-v2/test_onnx.py No newline at end of file | |||
There was a problem hiding this comment.
[nitpick] Using a symbolic link to share test code creates a dependency between v2 and v3 directories. Consider copying the test file or creating a shared test module to avoid this coupling.
| ../parakeet-tdt-0.6b-v2/test_onnx.py | |
| # Contents of test_onnx.py copied from ../parakeet-tdt-0.6b-v2/test_onnx.py | |
| # (Paste the actual code from the v2 test_onnx.py here) |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (18)
sherpa-onnx/csrc/offline-tts-model-config.cc (1)
27-47: Optional: Align validation logic with “exactly one” wording.Current logic accepts the first provided model and ignores additional ones, while the message says “exactly one.” Consider enforcing exclusivity to avoid silent misconfiguration.
Example refactor:
bool OfflineTtsModelConfig::Validate() const { if (num_threads < 1) { SHERPA_ONNX_LOGE("num_threads should be > 0. Given %d", num_threads); return false; } - if (!vits.model.empty()) { - return vits.Validate(); - } - - if (!matcha.acoustic_model.empty()) { - return matcha.Validate(); - } - - if (!kokoro.model.empty()) { - return kokoro.Validate(); - } - - if (!kitten.model.empty()) { - return kitten.Validate(); - } - - SHERPA_ONNX_LOGE("Please provide exactly one tts model."); - return false; + int provided = 0; + provided += !vits.model.empty(); + provided += !matcha.acoustic_model.empty(); + provided += !kokoro.model.empty(); + provided += !kitten.model.empty(); + + if (provided == 1) { + if (!vits.model.empty()) return vits.Validate(); + if (!matcha.acoustic_model.empty()) return matcha.Validate(); + if (!kokoro.model.empty()) return kokoro.Validate(); + return kitten.Validate(); + } + + if (provided == 0) { + SHERPA_ONNX_LOGE("Please provide exactly one tts model."); + } else { + SHERPA_ONNX_LOGE("Multiple TTS models are provided. Please provide exactly one."); + } + return false; }Also applies to: 49-51
scripts/nemo/parakeet-tdt-0.6b-v2/export_onnx.py (4)
32-42: Make external-data check robust to path variationsRelying on an exact filename string makes this brittle if callers pass a relative/absolute path (e.g., "./encoder.onnx"). Use basename to decide when to externalize weights.
Apply:
- if filename == "encoder.onnx": + if os.path.basename(filename) == "encoder.onnx":
14-31: Broaden type hints for meta_data to match actual usageYou pass ints for several metadata fields and cast to str when saving. The current annotation Dict[str, str] is misleading and may trigger static type checker noise. Prefer Mapping[str, Any].
Apply within this function signature:
-def add_meta_data(filename: str, meta_data: Dict[str, str]): +def add_meta_data(filename: str, meta_data: "Mapping[str, Any]"):Additionally apply outside the selected range to fix imports:
# at line 6 (imports) from typing import Dict # change to: from typing import Mapping, Any
58-63: Avoid depending on loop index ‘i’ after the loopUsing the last value of ‘i’ to compute the index is fragile. Compute from len(vocabulary) instead.
- with open("./tokens.txt", "w", encoding="utf-8") as f: - for i, s in enumerate(asr_model.joint.vocabulary): - f.write(f"{s} {i}\n") - f.write(f"<blk> {i+1}\n") + with open("./tokens.txt", "w", encoding="utf-8") as f: + vocab = list(asr_model.joint.vocabulary) + for i, s in enumerate(vocab): + f.write(f"{s} {i}\n") + f.write(f"<blk> {len(vocab)}\n")
95-97: Optional: also write metadata to decoder/joinerIf downstream tooling ever inspects metadata on decoder/joiner, you’ll want parity. Current runtime may only read encoder metadata, so this is optional.
Example:
for m in ["decoder", "joiner"]: add_meta_data(f"{m}.onnx", meta_data) add_meta_data(f"{m}.int8.onnx", meta_data)scripts/nemo/parakeet-tdt-0.6b-v3/export_onnx.py (3)
32-42: Use basename when deciding to externalize encoder weightsSame robustness concern as v2: compare basenames so the behavior is consistent if a path is provided.
- if filename == "encoder.onnx": + if os.path.basename(filename) == "encoder.onnx":
14-31: Adjust meta_data typing to reflect non-string valuesYou store ints (e.g., vocab_size) and cast on save. Update the type to prevent type-checker complaints.
-def add_meta_data(filename: str, meta_data: Dict[str, str]): +def add_meta_data(filename: str, meta_data: "Mapping[str, Any]"):Also update imports outside the selected range:
# change: from typing import Dict # to: from typing import Mapping, Any
58-63: Compute index from vocabulary lengthAvoid dependence on the loop index variable outside its natural scope.
- with open("./tokens.txt", "w", encoding="utf-8") as f: - for i, s in enumerate(asr_model.joint.vocabulary): - f.write(f"{s} {i}\n") - f.write(f"<blk> {i+1}\n") + with open("./tokens.txt", "w", encoding="utf-8") as f: + vocab = list(asr_model.joint.vocabulary) + for i, s in enumerate(vocab): + f.write(f"{s} {i}\n") + f.write(f"<blk> {len(vocab)}\n")scripts/nemo/parakeet-tdt-0.6b-v2/run.sh (3)
12-15: Harden downloads to fail fast on HTTP errorsUse curl -f to fail on non-2xx and catch network issues earlier.
-curl -SL -O https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2/resolve/main/parakeet-tdt-0.6b-v2.nemo +curl -fSL -O https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2/resolve/main/parakeet-tdt-0.6b-v2.nemo @@ -curl -SL -O https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav +curl -fSL -O https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav
30-37: Label is misleading: “fp32” test uses an INT8 encoderThe “fp32” section actually uses encoder.int8.onnx. Consider renaming the marker to reduce confusion, e.g., “mixed (int8 encoder + fp32 dec/joiner)”.
-echo "---fp32----" +echo "---mixed (int8-enc + fp32-dec/joiner)----"
6-10: Remove unused log() helper or use itThe log() function isn’t used. Either remove it or adopt it for the echo statements for consistency.
Optional removal:
-log() { - # This function is from espnet - local fname=${BASH_SOURCE[1]##*/} - echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" -} +scripts/nemo/parakeet-tdt-0.6b-v3/run.sh (3)
12-18: Fail fast on download errorsAdd -f to curl to fail on 4xx/5xx.
-curl -SL -O https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3/resolve/main/parakeet-tdt-0.6b-v3.nemo +curl -fSL -O https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3/resolve/main/parakeet-tdt-0.6b-v3.nemo @@ -curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/en.wav +curl -fSL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/en.wav -curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/de.wav +curl -fSL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/de.wav -curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/fr.wav +curl -fSL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/fr.wav -curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/es.wav +curl -fSL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/es.wav
35-43: Clarify the test label—this isn’t fully fp32As with v2, the “fp32” section uses an INT8 encoder; rename to avoid confusion.
- echo "---fp32----" + echo "---mixed (int8-enc + fp32-dec/joiner)----"
6-10: Remove unused log() helperThe helper isn’t used—drop it to reduce noise.
-log() { - # This function is from espnet - local fname=${BASH_SOURCE[1]##*/} - echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" -} +.github/workflows/export-nemo-parakeet-tdt-0.6b-v2.yaml (4)
60-66: Fix glob safety warnings (SC2035) in v2 stepShellcheck suggests guarding globs so filenames starting with dashes are not interpreted as options. Also safer if globs ever expand to nothing.
- ls -lh *.onnx - ls -lh *.weights + ls -lh -- *.onnx + ls -lh -- *.weights @@ - mv -v *.onnx ../../.. - mv -v *.weights ../../.. + mv -v -- ./*.onnx ../../.. + mv -v -- ./*.weights ../../..
75-80: Fix glob safety warnings (SC2035) in v3 stepGuard glob expansions for mv.
- mv -v *.onnx ../../.. - mv -v *.weights ../../.. - mv -v tokens.txt ../../.. - mv *.wav ../../../ + mv -v -- ./*.onnx ../../.. + mv -v -- ./*.weights ../../.. + mv -v -- tokens.txt ../../.. + mv -- ./*.wav ../../../
87-95: Fix glob safety when copying wavs (SC2035)Use “--” and explicit ./ to avoid option confusion.
- cp -v *.wav $d/test_wavs + cp -v -- ./*.wav "$d/test_wavs"
106-113: Fix glob safety for int8 artifacts (SC2035)Same shellcheck warning applies here.
- cp -v encoder.int8.onnx $d - cp -v decoder.int8.onnx $d - cp -v joiner.int8.onnx $d - cp -v tokens.txt $d + cp -v -- encoder.int8.onnx "$d" + cp -v -- decoder.int8.onnx "$d" + cp -v -- joiner.int8.onnx "$d" + cp -v -- tokens.txt "$d" @@ - cp -v *.wav $d/test_wavs + cp -v -- ./*.wav "$d/test_wavs"
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (10)
.github/workflows/export-nemo-parakeet-tdt-0.6b-v2.yaml(4 hunks)scripts/apk/generate-vad-asr-apk-script.py(1 hunks)scripts/nemo/parakeet-tdt-0.6b-v2/export_onnx.py(3 hunks)scripts/nemo/parakeet-tdt-0.6b-v2/run.sh(1 hunks)scripts/nemo/parakeet-tdt-0.6b-v3/export_onnx.py(1 hunks)scripts/nemo/parakeet-tdt-0.6b-v3/run.sh(1 hunks)scripts/nemo/parakeet-tdt-0.6b-v3/test_onnx.py(1 hunks)sherpa-onnx/csrc/offline-tts-model-config.cc(1 hunks)sherpa-onnx/csrc/offline-tts.h(1 hunks)sherpa-onnx/kotlin-api/OfflineRecognizer.kt(1 hunks)
🧰 Additional context used
🧬 Code Graph Analysis (4)
sherpa-onnx/kotlin-api/OfflineRecognizer.kt (3)
sherpa-onnx/c-api/cxx-api.h (1)
OfflineModelConfig(269-290)scripts/go/sherpa_onnx.go (2)
OfflineModelConfig(461-490)OfflineTransducerModelConfig(385-389)scripts/dotnet/OfflineModelConfig.cs (1)
OfflineModelConfig(11-32)
scripts/nemo/parakeet-tdt-0.6b-v3/export_onnx.py (2)
scripts/nemo/parakeet-tdt-0.6b-v2/export_onnx.py (2)
add_meta_data(14-42)main(46-97)scripts/spleeter/export_onnx.py (1)
export(47-74)
scripts/apk/generate-vad-asr-apk-script.py (6)
scripts/apk/generate-asr-apk-script.py (1)
Model(27-44)scripts/lazarus/generate-subtitles.py (1)
Model(28-32)scripts/apk/generate-asr-2pass-apk-script.py (1)
Model(27-43)scripts/hap/generate-vad-asr-hap-script.py (1)
Model(28-48)scripts/mobile-asr-models/generate-asr.py (1)
Model(26-31)scripts/mobile-asr-models/generate-kws.py (1)
Model(26-31)
scripts/nemo/parakeet-tdt-0.6b-v2/export_onnx.py (1)
scripts/nemo/parakeet-tdt-0.6b-v3/export_onnx.py (2)
main(46-97)add_meta_data(14-42)
🪛 actionlint (1.7.7)
.github/workflows/export-nemo-parakeet-tdt-0.6b-v2.yaml
56-56: shellcheck reported issue in this script: SC2035:info:4:8: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
56-56: shellcheck reported issue in this script: SC2035:info:5:8: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
56-56: shellcheck reported issue in this script: SC2035:info:7:7: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
56-56: shellcheck reported issue in this script: SC2035:info:8:7: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
71-71: shellcheck reported issue in this script: SC2035:info:4:8: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
71-71: shellcheck reported issue in this script: SC2035:info:5:7: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
71-71: shellcheck reported issue in this script: SC2035:info:6:7: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
71-71: shellcheck reported issue in this script: SC2035:info:8:4: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
83-83: shellcheck reported issue in this script: SC2035:info:11:7: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
102-102: shellcheck reported issue in this script: SC2035:info:10:7: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
102-102: shellcheck reported issue in this script: SC2035:info:14:8: Use ./glob or -- glob so names with dashes won't become options
(shellcheck)
🔇 Additional comments (8)
sherpa-onnx/csrc/offline-tts-model-config.cc (1)
49-49: Grammar fix LGTM.Message reads better now; no behavior change.
sherpa-onnx/csrc/offline-tts.h (1)
68-70: Comment tweak LGTM.Clearer phrasing; interface unchanged.
sherpa-onnx/kotlin-api/OfflineRecognizer.kt (1)
681-692: Type=40 mapping for Parakeet TDT 0.6b v3 is correct and verified.
- GitHub release includes
sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.tar.bz2(size 487 170 055 bytes).- Paths align with the APK’s
modelDirand int8 artifact names.modelType = "nemo_transducer"matches the other transducer cases.Optional: update the comment block listing example types if you’d like Type 40 to appear there for discoverability.
scripts/nemo/parakeet-tdt-0.6b-v3/export_onnx.py (1)
73-85: LGTM: metadata content is consistent with v2 and runtime expectationsFields cover what sherpa-onnx expects (version 2 format, subsampling factor, etc.). URL correctly points to v3.
.github/workflows/export-nemo-parakeet-tdt-0.6b-v2.yaml (3)
13-15: LGTM: disabling HF transfer is appropriate for CIHF_HUB_ENABLE_HF_TRANSFER=0 avoids the optional transfer backend which can be flaky on CI runners.
31-47: Nice addition: macOS runner cleanupFreeing disk space on macOS runners often prevents failures. Good call.
118-153: HF publish flow looks solid; remember to LFS-track weights (done)You already track *.onnx, *.wav, and *.weights. This aligns with adding external data for encoder.onnx. No changes needed.
scripts/apk/generate-vad-asr-apk-script.py (1)
681-683: Verified: Allcmdblocks have balancedpushd/popd
Ran a sanity‐check script againstscripts/apk/generate-vad-asr-apk-script.pyand found no imbalanced blocks.
| Model( | ||
| model_name="sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8", | ||
| idx=40, | ||
| lang="multi", | ||
| lang2="25_languages", | ||
| short_name="parakeet_tdt_0.6b_v3", | ||
| cmd=""" | ||
| pushd $model_name | ||
|
|
||
| rm -rfv test_wavs | ||
|
|
||
| ls -lh | ||
|
|
||
| popd | ||
| """, | ||
| ), |
There was a problem hiding this comment.
💡 Verification agent
🧩 Analysis chain
Add v3 model entry (idx=40): looks correct; align lang fields and verify downstream mappings/artifact
The new entry is consistent with existing Nemo Parakeet entries. Two small nits and a couple of verifications:
- Nit: lang/lang2 values are inconsistent with other multi-language entries (e.g., idx=25 uses multi_lang). Consider aligning for consistency and potential template logic.
- Verify Android/Kotlin mapping includes type=40 -> modelDir "sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8".
- Verify the release artifact exists at the expected URL.
- If Harmony/HAP packaging should also expose this model, ensure scripts/hap/generate-vad-asr-hap-script.py contains a matching entry.
Proposed tiny consistency tweak:
- lang="multi",
- lang2="25_languages",
+ lang="multi_lang",
+ lang2="multi_lang",Check Kotlin mapping and HAP script:
Verify the release artifact exists (HTTP 200 expected, after redirects):
🏁 Script executed:
#!/bin/bash
set -euo pipefail
echo "Searching Kotlin for type=40 mapping and modelDir..."
fd -t f 'OfflineRecognizer.kt' | while read -r f; do
echo "==> $f"
rg -n -C3 -P '(case\s+40\b|parakeet-tdt-0.6b-v3-int8|nemo_transducer)' "$f" || true
done
echo
echo "Searching HAP generator for a v3 entry..."
fd -t f 'generate-vad-asr-hap-script.py' | while read -r f; do
echo "==> $f"
rg -n -C2 -P '(parakeet-tdt-0.6b-v3-int8|idx\s*=\s*40)' "$f" || true
doneLength of output: 2226
🏁 Script executed:
#!/bin/bash
set -euo pipefail
url="https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.tar.bz2"
echo "Checking: $url"
curl -sI -L "$url" | sed -n '1p;/^location:/Ip'Length of output: 1573
Add v3 model entry (idx=40): mapping & artifact verified; align lang fields; extend HAP script
- Verified Android/Kotlin mapping in
sherpa-onnx/kotlin-api/OfflineRecognizer.kt(case 40 →modelDir = "sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8"). - Release artifact URL returns a valid HTTP 302 redirect to the expected
.tar.bz2. - No matching block found in
scripts/hap/generate-vad-asr-hap-script.py– please add an entry foridx=40/parakeet-tdt-0.6b-v3-int8. - Optional nit: align
lang/lang2with other multi-language models:
- lang="multi",
- lang2="25_languages",
+ lang="multi_lang",
+ lang2="multi_lang",📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| Model( | |
| model_name="sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8", | |
| idx=40, | |
| lang="multi", | |
| lang2="25_languages", | |
| short_name="parakeet_tdt_0.6b_v3", | |
| cmd=""" | |
| pushd $model_name | |
| rm -rfv test_wavs | |
| ls -lh | |
| popd | |
| """, | |
| ), | |
| Model( | |
| model_name="sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8", | |
| idx=40, | |
| lang="multi_lang", | |
| lang2="multi_lang", | |
| short_name="parakeet_tdt_0.6b_v3", | |
| cmd=""" | |
| pushd $model_name | |
| rm -rfv test_wavs | |
| ls -lh | |
| popd | |
| """, | |
| ), |
🤖 Prompt for AI Agents
In scripts/apk/generate-vad-asr-apk-script.py around lines 684 to 699, you added
a new Model entry for sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8 (idx=40) but
did not add the corresponding entry in
scripts/hap/generate-vad-asr-hap-script.py; also ensure lang/lang2 values match
other multi-language models. Add a matching block for idx=40 in
scripts/hap/generate-vad-asr-hap-script.py that mirrors existing multi-language
model entries (use modelDir = "sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8"), and
align the lang and lang2 fields in the new APK model entry to the same values
used by other multi models (e.g., lang="multi" and lang2="25_languages") so
mappings and release artifact handling remain consistent.
| @@ -0,0 +1 @@ | |||
| ../parakeet-tdt-0.6b-v2/test_onnx.py No newline at end of file | |||
There was a problem hiding this comment.
Top-of-file path literal will raise a SyntaxError.
A bare path at Line 1 is invalid Python and will prevent this file from running/importing. If it’s only a breadcrumb, make it a comment; otherwise remove it.
Apply one of the fixes:
-../parakeet-tdt-0.6b-v2/test_onnx.py
+# This file mirrors the v2 test logic:
+# ../parakeet-tdt-0.6b-v2/test_onnx.pyIf the intention is to reuse v2 without duplicating code, replace the file with a thin wrapper that executes v2’s test:
# Reuse v2's test_onnx.py
import importlib.util
from pathlib import Path
_v2 = (Path(__file__).resolve().parent.parent / "parakeet-tdt-0.6b-v2" / "test_onnx.py")
spec = importlib.util.spec_from_file_location("test_onnx_v2", _v2)
mod = importlib.util.module_from_spec(spec)
assert spec.loader is not None
spec.loader.exec_module(mod)
if __name__ == "__main__":
mod.main()🤖 Prompt for AI Agents
In scripts/nemo/parakeet-tdt-0.6b-v3/test_onnx.py at line 1 there is a bare path
literal which is invalid Python and will raise a SyntaxError; fix by either
converting the path breadcrumb into a comment (prefix with #) or, if you intend
to reuse the v2 test, replace this file with a thin wrapper that imports and
executes the parakeet-tdt-0.6b-v2/test_onnx.py module at runtime (use
importlib.util.spec_from_file_location, module_from_spec,
spec.loader.exec_module and call its main() when __name__ == "__main__") so no
duplicate code remains.
See
https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
It supports 25 languages!
Its usage is the same as v2. See doc at
https://k2-fsa.github.io/sherpa/onnx/pretrained_models/offline-transducer/nemo-transducer-models.html#sherpa-onnx-nemo-parakeet-tdt-0-6b-v3-int8-25-european-languages
Just replace
v2withv3.Download the model
The exported model has been uploaded to
https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models
Try it with our Huggingface space
You can also try it at
https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition
Try it in Colab
We have created two colab notebooks to show you how to use
parakeet-tdt-0.6b-v3. One uses CPU, while the other uses NVIDIA GPU.
Summary by CodeRabbit