Known issues and limitations for Watson Speech services

Watson Speech CR does not recover after PostgreSQL disk-full condition

Applies to: 5.4.0

Problem

After a storage full condition, the watson_speech custom resource (CR) does not recover and remains in an unhealthy state. The CR shows that it is waiting for the PostgreSQL cluster to enter a healthy state, but the PostgreSQL cluster fails to recover within the expected timeframe.

The watson_speech CR status shows:

watson_speech           WatsonSpeech             speech-cr                    2026-05-24T18:51:00Z  zen          5.4.0                 5.4.0               Watson Speech operator 12.0.0 build    25%         Waiting for Postgres cluster to be in a healthy state                2026-05-26T20:26:27.91567Z PostgreSQL cluster failed to enter a healthy state within 15 minutes  N/A         InProgress
edb_cp4d                CPDEdbService            cpd-edb-service              2026-05-24T18:24:26Z  z

Solution

Contact IBM Support for assistance.

Enabling Instana metrics causes degraded speech performance and increased latency

Applies to:5.4

Problem

When Instana® metrics are enabled, you might experience slow responses or intermittent timeouts for WebSocket‑based synthesize requests. Increased latency has also been observed across Watson Speech to Text (STT) and Watson Text to Speech (TTS) responses, indicating that Instana support for Watson Speech services is currently partial and can result in degraded performance.

Model upload jobs fail when Amazon S3 bucket name exceeds length limits

Applies to: 5.4

Problem

Speech model upload jobs can fail with NoSuchBucket errors during execution, and the associated job pods remain in the Error state.

oc get pods -n ibm-software-hub-operands -w | grep -iE 'model|upload'
speech-cr-stt-models-ce8c5d75ee74b3e0ead636d75b953a8feecda2dvqv   5/6     Error   0   60s

Pod logs show that the upload process attempts to create or access an Amazon S3 bucket with a generated name that exceeds the S3 maximum bucket name length of 63 characters. When the bucket name exceeds this limit, Amazon S3 rejects the request and the job fails with errors such as

botocore.errorfactory.NoSuchBucket: An error occurred (NoSuchBucket) when calling the PutObject operation: The specified bucket does not exist.

By default, the generated bucket name follows the pattern speech-service-base-models-ibm-<release-name>-<namespace>. If the <namespace> value is longer than 21 characters, the resulting bucket name exceeds the Amazon S3 naming limit and cannot be created.

Solution

Edit the Speech custom resource (CR) and explicitly override the bucketSuffix value to ensure that the final bucket name stays within the 63‑character limit.

spec:
  global:
    datastores:
      s3:
        bucketSuffix: "ibm-{{ releaseName }}-{{ speechNamespace }}"

cpd-cli export-import command fails when EXPORT_NAME exceeds 14 characters

Applies to: 5.4

Problem

EXPORT_NAME used in the cpd-cli export-import command can contain a maximum of 14 characters. The command fails and returns the following error when EXPORT_NAME contains more than 14 characters.

Error: ConfigMap "cpd-ex-5.2.2-stt-export-speech-to-text-aux-1763738353694737-ei-cm" is invalid: metadata.labels: Invalid value: "cpd-ex-5.2.2-stt-export-speech-to-text-aux-1763738353694737-ei-cm": must be no more than 63 characters
[ERROR] 2025-11-21T09:02:12.621725Z RunPluginCommand:Execution error:  exit status 1

Solution

Define EXPORT_NAME using no more than 14 characters.

Language detection with hybrid models and rnn-t models ignores lid_confidence parameter

Applies to: 5.4

Problem

When you use the detect_language endpoint with a hybrid model or a rnn-t model (for example, detect_language?model=en-US_BroadbandModel), the lid_confidence query parameter is ignored. In this case, the request is processed as a standard recognize/ request, which counts against the client’s license quota. This behavior differs from the intended use of detect_language as a free API.

Solution

Use only Large Speech Models (LSMs) for language detection.

Language model customization fails on en-US_Telephony or en-US_Multimedia models with 500 Internal Server Error

Applies to: 5.4

Problem

When you train language model customizations by using en-US_Telephony or en-US_Multimedia models, it might fail with a similar error.

DefaultCPUAllocator: can't allocate memory: you tried to allocate X bytes. Error code 12 (Cannot allocate memory)

Solution

Use the English Large Speech Model (en-US). If the model does not work, contact IBM Support.